Model save

README.md

@@ -0,0 +1,55 @@
+---
+tags:
+- trl
+- sft
+- generated_from_trainer
+model-index:
+- name: Phi-3-mini-4k-instruct-full-finetuned-v01
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# Phi-3-mini-4k-instruct-full-finetuned-v01
+
+This model was trained from scratch on an unknown dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.4590
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 3e-05
+- train_batch_size: 2
+- eval_batch_size: 2
+- seed: 0
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.2
+- num_epochs: 1
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 4.43.3
+- Pytorch 2.2.2+cu121
+- Datasets 2.20.0
+- Tokenizers 0.19.1

all_results.json

@@ -0,0 +1,13 @@
+{
+    "epoch": 1.0,
+    "eval_loss": 0.4590456783771515,
+    "eval_runtime": 233.2276,
+    "eval_samples": 558,
+    "eval_samples_per_second": 2.393,
+    "eval_steps_per_second": 1.196,
+    "total_flos": 5.485972481640161e+17,
+    "train_loss": 0.4824531834622566,
+    "train_runtime": 26326.043,
+    "train_samples_per_second": 0.685,
+    "train_steps_per_second": 0.342
+}

configuration_phi3.py

@@ -0,0 +1,227 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Phi-3 model configuration"""
+
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/Phi-3-mini-4k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json",
+    "microsoft/Phi-3-mini-128k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json",
+}
+
+
+class Phi3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the
+    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32064):
+            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Phi3Model`].
+        hidden_size (`int`, *optional*, defaults to 3072):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            Dropout probability for mlp outputs.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio after computing the attention scores.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        original_max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model was trained with. This is used to determine the size of the
+            original RoPE embeddings when using long scaling.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon value used for the RMSNorm.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`dict`, *optional*):
+            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
+            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be `longrope` and 
+            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size 
+            divided by the number of attention heads divided by 2.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 32000):
+            The id of the "end-of-sequence" token.
+        pad_token_id (`int`, *optional*, defaults to 32000):
+            The id of the padding token.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If `None`, no sliding window is applied.
+
+    Example:
+
+    ```python
+    >>> from transformers import Phi3Model, Phi3Config
+
+    >>> # Initializing a Phi-3 style configuration
+    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
+
+    >>> # Initializing a model from the configuration
+    >>> model = Phi3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "phi3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32064,
+        hidden_size=3072,
+        intermediate_size=8192,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attention_dropout=0.0,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        original_max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        bos_token_id=1,
+        eos_token_id=32000,
+        pad_token_id=32000,
+        sliding_window=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attention_dropout = attention_dropout
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self._rope_scaling_adjustment()
+        self._rope_scaling_validation()
+        self.sliding_window = sliding_window
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    def _rope_scaling_adjustment(self):
+        """
+        Adjust the `type` of the `rope_scaling` configuration for backward compatibility.
+        """
+        if self.rope_scaling is None:
+            return
+
+        rope_scaling_type = self.rope_scaling.get("type", None)
+
+        # For backward compatibility if previous version used "su" or "yarn"
+        if rope_scaling_type is not None and rope_scaling_type in ["su", "yarn"]:
+            self.rope_scaling["type"] = "longrope"
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
+            raise ValueError(
+                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
+                f"got {self.rope_scaling}"
+            )
+        rope_scaling_type = self.rope_scaling.get("type", None)
+        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
+        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
+        if rope_scaling_type is None or rope_scaling_type not in ["longrope"]:
+            raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], got {rope_scaling_type}")
+        if not (
+            isinstance(rope_scaling_short_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
+            )
+        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
+            )
+        if not (
+            isinstance(rope_scaling_long_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
+            )
+        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
+            )

eval_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.0,
+    "eval_loss": 0.4590456783771515,
+    "eval_runtime": 233.2276,
+    "eval_samples": 558,
+    "eval_samples_per_second": 2.393,
+    "eval_steps_per_second": 1.196
+}

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    32000,
+    32001,
+    32007
+  ],
+  "pad_token_id": 32000,
+  "transformers_version": "4.43.3"
+}

model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7d2542c0809c7e2d208b889dca2a30bb66de676de582cfb4bb9c956d989856fe
+oid sha256:ab568a422e189be32ef82332bac5922fa40d18cf1acde2ed49a039938bf304be
 size 4972489328

model-00002-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:138c87db7f21c3f52a7499fdf9f967ceac3f2af29521ef96d11c9e863a65912d
+oid sha256:afd166e96177532d652009e1487f4abd1d47898d57be8bcdaea86e2a155cb388
 size 2669692552

modeling_phi3.py

@@ -0,0 +1,1563 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" PyTorch Phi-3 model."""
+
+import inspect
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_phi3 import Phi3Config
+
+
+logger = logging.get_logger(__name__)
+
+# Transformers scans dependencies in the modeling file, causing issues on conditional loading. The regex only ignores try/catch blocks, but not if statements
+# if is_flash_attn_2_available():
+_flash_supports_window_size = False
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+except ImportError as error:
+    logger.warning(
+        f"`flash-attention` package not found, consider installing for better performance: {error}."
+    )
+    if not _flash_supports_window_size:
+        logger.warning(
+            "Current `flash-attention` does not support `window_size`. Either upgrade or use `attn_implementation='eager'`."
+        )
+
+_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
+_CONFIG_FOR_DOC = "Phi3Config"
+
+PHI3_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/Phi-3-mini-4k-instruct",
+    "microsoft/Phi-3-mini-128k-instruct",
+    # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3
+]
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
+class Phi3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Phi3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
+class Phi3RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        self.register_buffer("inv_freq", None, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if self.inv_freq is None:
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
+            )
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3LongRoPEScaledRotaryEmbedding(Phi3RotaryEmbedding):
+    def __init__(self, dim, config, device=None):
+        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
+
+        self.short_factor = config.rope_scaling["short_factor"]
+        self.long_factor = config.rope_scaling["long_factor"]
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.original_max_position_embeddings:
+            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
+        else:
+            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
+
+        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
+        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+
+            scale = self.max_position_embeddings / self.original_max_position_embeddings
+            if scale <= 1.0:
+                scaling_factor = 1.0
+            else:
+                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
+
+            cos = emb.cos() * scaling_factor
+            sin = emb.sin() * scaling_factor
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+        self.activation_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
+
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
+
+        return self.down_proj(up_states)
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Phi3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = config.rope_scaling
+        self.is_causal = True
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
+        self._init_rope()
+
+    def _init_rope(self):
+        if self.rope_scaling is None:
+            self.rotary_emb = Phi3RotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+        else:
+            scaling_type = self.config.rope_scaling["type"]
+            if scaling_type == "longrope":
+                self.rotary_emb = Phi3LongRoPEScaledRotaryEmbedding(self.head_dim, self.config)
+            else:
+                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class Phi3FlashAttention2(Phi3Attention):
+    """
+    Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # Phi3FlashAttention2 attention does not support output_attentions
+
+        if not _flash_supports_window_size:
+            logger.warning_once(
+                "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
+            )
+            raise ValueError("The current flash attention version does not support sliding window attention.")
+
+        output_attentions = False
+
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop("padding_mask")
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        # Because the input can be padded, the absolute sequence length depends on the max position id.
+        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        use_sliding_windows = (
+            _flash_supports_window_size
+            and getattr(self.config, "sliding_window", None) is not None
+            and kv_seq_len > self.config.sliding_window
+        )
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
+                slicing_tokens = 1 - self.config.sliding_window
+
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+                        f" {past_key.shape}"
+                    )
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_dropout = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32.
+
+        if query_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.qkv_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=attn_dropout,
+            use_sliding_windows=use_sliding_windows,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+        use_sliding_windows=False,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            use_sliding_windows (`bool`, *optional*):
+                Whether to activate sliding window attention.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            if not use_sliding_windows:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            if not use_sliding_windows:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+        return attn_output
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+        # On the first iteration we need to properly re-create the padding mask
+        # by slicing it on the proper place
+        if kv_seq_len != attention_mask.shape[-1]:
+            attention_mask_num_tokens = attention_mask.shape[-1]
+            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
+# TODO @Arthur no longer copied from LLama after static cache
+class Phi3SdpaAttention(Phi3Attention):
+    """
+    Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from Phi3Attention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+PHI3_ATTENTION_CLASSES = {
+    "eager": Phi3Attention,
+    "flash_attention_2": Phi3FlashAttention2,
+    "sdpa": Phi3SdpaAttention,
+}
+
+
+class Phi3DecoderLayer(nn.Module):
+    def __init__(self, config: Phi3Config, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+
+        self.mlp = Phi3MLP(config)
+        self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
+        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
+                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        attn_outputs, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+
+        hidden_states = residual + self.resid_attn_dropout(attn_outputs)
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+PHI3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Phi3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3PreTrainedModel(PreTrainedModel):
+    config_class = Phi3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Phi3DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+    _supports_cache_class = True
+
+    _version = "0.0.5"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+PHI3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3Model(Phi3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
+
+    Args:
+        config: Phi3Config
+    """
+
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.embed_dropout = nn.Dropout(config.embd_pdrop)
+        self.layers = nn.ModuleList(
+            [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        past_key_values_length = 0
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
+            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            if is_padding_right:
+                raise ValueError(
+                    "You are attempting to perform batched generation with padding_side='right'"
+                    " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
+                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                )
+
+        if self._attn_implementation == "flash_attention_2":
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+                sliding_window=self.config.sliding_window,
+            )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class Phi3ForCausalLM(Phi3PreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Phi3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
+    def get_decoder(self):
+        return self.model
+
+    # Ignore copy
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Phi3ForCausalLM
+
+        >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+
+        >>> prompt = "This is an example script ."
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The [`Phi3Model`] with a sequence classification head on top (linear layer).
+
+    [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
+class Phi3ForSequenceClassification(Phi3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Phi3Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = model_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + model_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=model_outputs.past_key_values,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
+class Phi3ForTokenClassification(Phi3PreTrainedModel):
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.model = Phi3Model(config)
+        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = model_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            batch_size, seq_length = labels.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (logits,) + model_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )

tokenizer_config.json

@@ -123,7 +123,7 @@
   "legacy": false,
   "model_max_length": 4096,
   "pad_token": "<unk>",
-  "padding_side": "right",
+  "padding_side": "left",
   "sp_model_kwargs": {},
   "tokenizer_class": "LlamaTokenizer",
   "unk_token": "<unk>",

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.0,
+    "total_flos": 5.485972481640161e+17,
+    "train_loss": 0.4824531834622566,
+    "train_runtime": 26326.043,
+    "train_samples_per_second": 0.685,
+    "train_steps_per_second": 0.342
+}

trainer_state.json

@@ -0,0 +1,3192 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 500,
+  "global_step": 9013,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0022190169754798626,
+      "grad_norm": 7.625,
+      "learning_rate": 3.3277870216306157e-07,
+      "loss": 1.2617,
+      "step": 20
+    },
+    {
+      "epoch": 0.004438033950959725,
+      "grad_norm": 5.5625,
+      "learning_rate": 6.655574043261231e-07,
+      "loss": 1.2121,
+      "step": 40
+    },
+    {
+      "epoch": 0.006657050926439588,
+      "grad_norm": 4.78125,
+      "learning_rate": 9.983361064891848e-07,
+      "loss": 1.2737,
+      "step": 60
+    },
+    {
+      "epoch": 0.00887606790191945,
+      "grad_norm": 5.46875,
+      "learning_rate": 1.3311148086522463e-06,
+      "loss": 1.1634,
+      "step": 80
+    },
+    {
+      "epoch": 0.011095084877399313,
+      "grad_norm": 6.46875,
+      "learning_rate": 1.6638935108153078e-06,
+      "loss": 1.1582,
+      "step": 100
+    },
+    {
+      "epoch": 0.013314101852879175,
+      "grad_norm": 6.09375,
+      "learning_rate": 1.9966722129783695e-06,
+      "loss": 1.1071,
+      "step": 120
+    },
+    {
+      "epoch": 0.015533118828359036,
+      "grad_norm": 5.5625,
+      "learning_rate": 2.329450915141431e-06,
+      "loss": 1.0944,
+      "step": 140
+    },
+    {
+      "epoch": 0.0177521358038389,
+      "grad_norm": 3.3125,
+      "learning_rate": 2.6622296173044925e-06,
+      "loss": 0.9622,
+      "step": 160
+    },
+    {
+      "epoch": 0.01997115277931876,
+      "grad_norm": 5.65625,
+      "learning_rate": 2.995008319467554e-06,
+      "loss": 0.7219,
+      "step": 180
+    },
+    {
+      "epoch": 0.022190169754798626,
+      "grad_norm": 3.234375,
+      "learning_rate": 3.3277870216306156e-06,
+      "loss": 0.6194,
+      "step": 200
+    },
+    {
+      "epoch": 0.024409186730278486,
+      "grad_norm": 3.359375,
+      "learning_rate": 3.6605657237936775e-06,
+      "loss": 0.5223,
+      "step": 220
+    },
+    {
+      "epoch": 0.02662820370575835,
+      "grad_norm": 2.234375,
+      "learning_rate": 3.993344425956739e-06,
+      "loss": 0.5673,
+      "step": 240
+    },
+    {
+      "epoch": 0.02884722068123821,
+      "grad_norm": 2.109375,
+      "learning_rate": 4.326123128119801e-06,
+      "loss": 0.5215,
+      "step": 260
+    },
+    {
+      "epoch": 0.031066237656718072,
+      "grad_norm": 2.765625,
+      "learning_rate": 4.658901830282862e-06,
+      "loss": 0.5166,
+      "step": 280
+    },
+    {
+      "epoch": 0.03328525463219793,
+      "grad_norm": 2.0625,
+      "learning_rate": 4.991680532445923e-06,
+      "loss": 0.5023,
+      "step": 300
+    },
+    {
+      "epoch": 0.0355042716076778,
+      "grad_norm": 2.0625,
+      "learning_rate": 5.324459234608985e-06,
+      "loss": 0.4611,
+      "step": 320
+    },
+    {
+      "epoch": 0.03772328858315766,
+      "grad_norm": 2.09375,
+      "learning_rate": 5.657237936772047e-06,
+      "loss": 0.5203,
+      "step": 340
+    },
+    {
+      "epoch": 0.03994230555863752,
+      "grad_norm": 2.34375,
+      "learning_rate": 5.990016638935108e-06,
+      "loss": 0.5498,
+      "step": 360
+    },
+    {
+      "epoch": 0.04216132253411738,
+      "grad_norm": 2.03125,
+      "learning_rate": 6.32279534109817e-06,
+      "loss": 0.5597,
+      "step": 380
+    },
+    {
+      "epoch": 0.04438033950959725,
+      "grad_norm": 2.453125,
+      "learning_rate": 6.655574043261231e-06,
+      "loss": 0.5133,
+      "step": 400
+    },
+    {
+      "epoch": 0.04659935648507711,
+      "grad_norm": 2.15625,
+      "learning_rate": 6.988352745424292e-06,
+      "loss": 0.4843,
+      "step": 420
+    },
+    {
+      "epoch": 0.04881837346055697,
+      "grad_norm": 2.265625,
+      "learning_rate": 7.321131447587355e-06,
+      "loss": 0.5358,
+      "step": 440
+    },
+    {
+      "epoch": 0.051037390436036834,
+      "grad_norm": 2.6875,
+      "learning_rate": 7.653910149750416e-06,
+      "loss": 0.3936,
+      "step": 460
+    },
+    {
+      "epoch": 0.0532564074115167,
+      "grad_norm": 2.125,
+      "learning_rate": 7.986688851913478e-06,
+      "loss": 0.5245,
+      "step": 480
+    },
+    {
+      "epoch": 0.05547542438699656,
+      "grad_norm": 2.796875,
+      "learning_rate": 8.319467554076538e-06,
+      "loss": 0.5094,
+      "step": 500
+    },
+    {
+      "epoch": 0.05769444136247642,
+      "grad_norm": 1.84375,
+      "learning_rate": 8.652246256239602e-06,
+      "loss": 0.4476,
+      "step": 520
+    },
+    {
+      "epoch": 0.059913458337956284,
+      "grad_norm": 2.09375,
+      "learning_rate": 8.985024958402662e-06,
+      "loss": 0.4836,
+      "step": 540
+    },
+    {
+      "epoch": 0.062132475313436145,
+      "grad_norm": 1.6328125,
+      "learning_rate": 9.317803660565724e-06,
+      "loss": 0.4805,
+      "step": 560
+    },
+    {
+      "epoch": 0.06435149228891601,
+      "grad_norm": 1.8828125,
+      "learning_rate": 9.650582362728786e-06,
+      "loss": 0.425,
+      "step": 580
+    },
+    {
+      "epoch": 0.06657050926439587,
+      "grad_norm": 2.6875,
+      "learning_rate": 9.983361064891846e-06,
+      "loss": 0.5223,
+      "step": 600
+    },
+    {
+      "epoch": 0.06878952623987573,
+      "grad_norm": 2.140625,
+      "learning_rate": 1.031613976705491e-05,
+      "loss": 0.5,
+      "step": 620
+    },
+    {
+      "epoch": 0.0710085432153556,
+      "grad_norm": 2.203125,
+      "learning_rate": 1.064891846921797e-05,
+      "loss": 0.5271,
+      "step": 640
+    },
+    {
+      "epoch": 0.07322756019083546,
+      "grad_norm": 1.9765625,
+      "learning_rate": 1.0981697171381032e-05,
+      "loss": 0.4981,
+      "step": 660
+    },
+    {
+      "epoch": 0.07544657716631532,
+      "grad_norm": 1.2421875,
+      "learning_rate": 1.1314475873544094e-05,
+      "loss": 0.5133,
+      "step": 680
+    },
+    {
+      "epoch": 0.07766559414179519,
+      "grad_norm": 2.515625,
+      "learning_rate": 1.1647254575707154e-05,
+      "loss": 0.4693,
+      "step": 700
+    },
+    {
+      "epoch": 0.07988461111727505,
+      "grad_norm": 2.03125,
+      "learning_rate": 1.1980033277870216e-05,
+      "loss": 0.4844,
+      "step": 720
+    },
+    {
+      "epoch": 0.08210362809275491,
+      "grad_norm": 1.578125,
+      "learning_rate": 1.2312811980033278e-05,
+      "loss": 0.4943,
+      "step": 740
+    },
+    {
+      "epoch": 0.08432264506823477,
+      "grad_norm": 2.75,
+      "learning_rate": 1.264559068219634e-05,
+      "loss": 0.5,
+      "step": 760
+    },
+    {
+      "epoch": 0.08654166204371463,
+      "grad_norm": 1.46875,
+      "learning_rate": 1.2978369384359402e-05,
+      "loss": 0.4318,
+      "step": 780
+    },
+    {
+      "epoch": 0.0887606790191945,
+      "grad_norm": 2.484375,
+      "learning_rate": 1.3311148086522462e-05,
+      "loss": 0.4745,
+      "step": 800
+    },
+    {
+      "epoch": 0.09097969599467436,
+      "grad_norm": 2.28125,
+      "learning_rate": 1.3643926788685524e-05,
+      "loss": 0.5459,
+      "step": 820
+    },
+    {
+      "epoch": 0.09319871297015422,
+      "grad_norm": 2.078125,
+      "learning_rate": 1.3976705490848584e-05,
+      "loss": 0.5021,
+      "step": 840
+    },
+    {
+      "epoch": 0.09541772994563408,
+      "grad_norm": 0.97265625,
+      "learning_rate": 1.4309484193011648e-05,
+      "loss": 0.4271,
+      "step": 860
+    },
+    {
+      "epoch": 0.09763674692111395,
+      "grad_norm": 1.828125,
+      "learning_rate": 1.464226289517471e-05,
+      "loss": 0.4924,
+      "step": 880
+    },
+    {
+      "epoch": 0.09985576389659381,
+      "grad_norm": 2.25,
+      "learning_rate": 1.497504159733777e-05,
+      "loss": 0.514,
+      "step": 900
+    },
+    {
+      "epoch": 0.10207478087207367,
+      "grad_norm": 1.578125,
+      "learning_rate": 1.5307820299500832e-05,
+      "loss": 0.4569,
+      "step": 920
+    },
+    {
+      "epoch": 0.10429379784755353,
+      "grad_norm": 1.4140625,
+      "learning_rate": 1.5640599001663892e-05,
+      "loss": 0.4751,
+      "step": 940
+    },
+    {
+      "epoch": 0.1065128148230334,
+      "grad_norm": 2.1875,
+      "learning_rate": 1.5973377703826956e-05,
+      "loss": 0.4719,
+      "step": 960
+    },
+    {
+      "epoch": 0.10873183179851326,
+      "grad_norm": 1.921875,
+      "learning_rate": 1.6306156405990016e-05,
+      "loss": 0.4455,
+      "step": 980
+    },
+    {
+      "epoch": 0.11095084877399312,
+      "grad_norm": 1.7421875,
+      "learning_rate": 1.6638935108153077e-05,
+      "loss": 0.5068,
+      "step": 1000
+    },
+    {
+      "epoch": 0.11316986574947298,
+      "grad_norm": 1.96875,
+      "learning_rate": 1.697171381031614e-05,
+      "loss": 0.5076,
+      "step": 1020
+    },
+    {
+      "epoch": 0.11538888272495285,
+      "grad_norm": 2.0625,
+      "learning_rate": 1.7304492512479204e-05,
+      "loss": 0.449,
+      "step": 1040
+    },
+    {
+      "epoch": 0.11760789970043271,
+      "grad_norm": 2.859375,
+      "learning_rate": 1.7637271214642264e-05,
+      "loss": 0.4298,
+      "step": 1060
+    },
+    {
+      "epoch": 0.11982691667591257,
+      "grad_norm": 2.0625,
+      "learning_rate": 1.7970049916805324e-05,
+      "loss": 0.5298,
+      "step": 1080
+    },
+    {
+      "epoch": 0.12204593365139244,
+      "grad_norm": 1.90625,
+      "learning_rate": 1.8302828618968388e-05,
+      "loss": 0.5398,
+      "step": 1100
+    },
+    {
+      "epoch": 0.12426495062687229,
+      "grad_norm": 1.40625,
+      "learning_rate": 1.8635607321131448e-05,
+      "loss": 0.5533,
+      "step": 1120
+    },
+    {
+      "epoch": 0.12648396760235217,
+      "grad_norm": 2.21875,
+      "learning_rate": 1.896838602329451e-05,
+      "loss": 0.422,
+      "step": 1140
+    },
+    {
+      "epoch": 0.12870298457783202,
+      "grad_norm": 1.96875,
+      "learning_rate": 1.9301164725457572e-05,
+      "loss": 0.4807,
+      "step": 1160
+    },
+    {
+      "epoch": 0.13092200155331188,
+      "grad_norm": 1.484375,
+      "learning_rate": 1.9633943427620632e-05,
+      "loss": 0.5209,
+      "step": 1180
+    },
+    {
+      "epoch": 0.13314101852879173,
+      "grad_norm": 1.9765625,
+      "learning_rate": 1.9966722129783693e-05,
+      "loss": 0.4522,
+      "step": 1200
+    },
+    {
+      "epoch": 0.13536003550427161,
+      "grad_norm": 1.3828125,
+      "learning_rate": 2.0299500831946756e-05,
+      "loss": 0.4684,
+      "step": 1220
+    },
+    {
+      "epoch": 0.13757905247975147,
+      "grad_norm": 1.59375,
+      "learning_rate": 2.063227953410982e-05,
+      "loss": 0.3982,
+      "step": 1240
+    },
+    {
+      "epoch": 0.13979806945523132,
+      "grad_norm": 1.7734375,
+      "learning_rate": 2.096505823627288e-05,
+      "loss": 0.5518,
+      "step": 1260
+    },
+    {
+      "epoch": 0.1420170864307112,
+      "grad_norm": 1.6875,
+      "learning_rate": 2.129783693843594e-05,
+      "loss": 0.5107,
+      "step": 1280
+    },
+    {
+      "epoch": 0.14423610340619106,
+      "grad_norm": 1.453125,
+      "learning_rate": 2.1630615640599004e-05,
+      "loss": 0.4422,
+      "step": 1300
+    },
+    {
+      "epoch": 0.1464551203816709,
+      "grad_norm": 2.015625,
+      "learning_rate": 2.1963394342762064e-05,
+      "loss": 0.5578,
+      "step": 1320
+    },
+    {
+      "epoch": 0.1486741373571508,
+      "grad_norm": 1.4921875,
+      "learning_rate": 2.2296173044925124e-05,
+      "loss": 0.4516,
+      "step": 1340
+    },
+    {
+      "epoch": 0.15089315433263065,
+      "grad_norm": 2.15625,
+      "learning_rate": 2.2628951747088188e-05,
+      "loss": 0.3979,
+      "step": 1360
+    },
+    {
+      "epoch": 0.1531121713081105,
+      "grad_norm": 1.4140625,
+      "learning_rate": 2.296173044925125e-05,
+      "loss": 0.4895,
+      "step": 1380
+    },
+    {
+      "epoch": 0.15533118828359038,
+      "grad_norm": 1.453125,
+      "learning_rate": 2.329450915141431e-05,
+      "loss": 0.4505,
+      "step": 1400
+    },
+    {
+      "epoch": 0.15755020525907024,
+      "grad_norm": 2.265625,
+      "learning_rate": 2.3627287853577372e-05,
+      "loss": 0.5008,
+      "step": 1420
+    },
+    {
+      "epoch": 0.1597692222345501,
+      "grad_norm": 1.5390625,
+      "learning_rate": 2.3960066555740432e-05,
+      "loss": 0.523,
+      "step": 1440
+    },
+    {
+      "epoch": 0.16198823921002994,
+      "grad_norm": 1.5546875,
+      "learning_rate": 2.4292845257903493e-05,
+      "loss": 0.4813,
+      "step": 1460
+    },
+    {
+      "epoch": 0.16420725618550983,
+      "grad_norm": 1.46875,
+      "learning_rate": 2.4625623960066556e-05,
+      "loss": 0.4995,
+      "step": 1480
+    },
+    {
+      "epoch": 0.16642627316098968,
+      "grad_norm": 1.8359375,
+      "learning_rate": 2.495840266222962e-05,
+      "loss": 0.474,
+      "step": 1500
+    },
+    {
+      "epoch": 0.16864529013646953,
+      "grad_norm": 1.6328125,
+      "learning_rate": 2.529118136439268e-05,
+      "loss": 0.4201,
+      "step": 1520
+    },
+    {
+      "epoch": 0.17086430711194942,
+      "grad_norm": 2.15625,
+      "learning_rate": 2.562396006655574e-05,
+      "loss": 0.5332,
+      "step": 1540
+    },
+    {
+      "epoch": 0.17308332408742927,
+      "grad_norm": 2.53125,
+      "learning_rate": 2.5956738768718804e-05,
+      "loss": 0.5105,
+      "step": 1560
+    },
+    {
+      "epoch": 0.17530234106290912,
+      "grad_norm": 1.734375,
+      "learning_rate": 2.6289517470881864e-05,
+      "loss": 0.4679,
+      "step": 1580
+    },
+    {
+      "epoch": 0.177521358038389,
+      "grad_norm": 1.4453125,
+      "learning_rate": 2.6622296173044925e-05,
+      "loss": 0.4276,
+      "step": 1600
+    },
+    {
+      "epoch": 0.17974037501386886,
+      "grad_norm": 1.9609375,
+      "learning_rate": 2.6955074875207988e-05,
+      "loss": 0.4462,
+      "step": 1620
+    },
+    {
+      "epoch": 0.1819593919893487,
+      "grad_norm": 1.734375,
+      "learning_rate": 2.728785357737105e-05,
+      "loss": 0.4534,
+      "step": 1640
+    },
+    {
+      "epoch": 0.1841784089648286,
+      "grad_norm": 1.5703125,
+      "learning_rate": 2.762063227953411e-05,
+      "loss": 0.4699,
+      "step": 1660
+    },
+    {
+      "epoch": 0.18639742594030845,
+      "grad_norm": 1.828125,
+      "learning_rate": 2.795341098169717e-05,
+      "loss": 0.483,
+      "step": 1680
+    },
+    {
+      "epoch": 0.1886164429157883,
+      "grad_norm": 2.765625,
+      "learning_rate": 2.8286189683860236e-05,
+      "loss": 0.4776,
+      "step": 1700
+    },
+    {
+      "epoch": 0.19083545989126816,
+      "grad_norm": 2.0,
+      "learning_rate": 2.8618968386023296e-05,
+      "loss": 0.4361,
+      "step": 1720
+    },
+    {
+      "epoch": 0.19305447686674804,
+      "grad_norm": 1.640625,
+      "learning_rate": 2.8951747088186356e-05,
+      "loss": 0.4566,
+      "step": 1740
+    },
+    {
+      "epoch": 0.1952734938422279,
+      "grad_norm": 1.640625,
+      "learning_rate": 2.928452579034942e-05,
+      "loss": 0.4609,
+      "step": 1760
+    },
+    {
+      "epoch": 0.19749251081770774,
+      "grad_norm": 1.9140625,
+      "learning_rate": 2.961730449251248e-05,
+      "loss": 0.5019,
+      "step": 1780
+    },
+    {
+      "epoch": 0.19971152779318763,
+      "grad_norm": 2.125,
+      "learning_rate": 2.995008319467554e-05,
+      "loss": 0.4844,
+      "step": 1800
+    },
+    {
+      "epoch": 0.20193054476866748,
+      "grad_norm": 2.015625,
+      "learning_rate": 2.999958848436878e-05,
+      "loss": 0.4974,
+      "step": 1820
+    },
+    {
+      "epoch": 0.20414956174414733,
+      "grad_norm": 1.828125,
+      "learning_rate": 2.9998050673796383e-05,
+      "loss": 0.4591,
+      "step": 1840
+    },
+    {
+      "epoch": 0.20636857871962722,
+      "grad_norm": 1.8671875,
+      "learning_rate": 2.999537386964595e-05,
+      "loss": 0.4319,
+      "step": 1860
+    },
+    {
+      "epoch": 0.20858759569510707,
+      "grad_norm": 1.953125,
+      "learning_rate": 2.9991558275201416e-05,
+      "loss": 0.4425,
+      "step": 1880
+    },
+    {
+      "epoch": 0.21080661267058692,
+      "grad_norm": 1.703125,
+      "learning_rate": 2.998660418022961e-05,
+      "loss": 0.4729,
+      "step": 1900
+    },
+    {
+      "epoch": 0.2130256296460668,
+      "grad_norm": 1.6484375,
+      "learning_rate": 2.9980511960958247e-05,
+      "loss": 0.4049,
+      "step": 1920
+    },
+    {
+      "epoch": 0.21524464662154666,
+      "grad_norm": 1.671875,
+      "learning_rate": 2.9973282080047365e-05,
+      "loss": 0.4646,
+      "step": 1940
+    },
+    {
+      "epoch": 0.2174636635970265,
+      "grad_norm": 1.6875,
+      "learning_rate": 2.996491508655417e-05,
+      "loss": 0.4725,
+      "step": 1960
+    },
+    {
+      "epoch": 0.21968268057250637,
+      "grad_norm": 1.953125,
+      "learning_rate": 2.995541161589137e-05,
+      "loss": 0.4448,
+      "step": 1980
+    },
+    {
+      "epoch": 0.22190169754798625,
+      "grad_norm": 1.3046875,
+      "learning_rate": 2.994477238977889e-05,
+      "loss": 0.4673,
+      "step": 2000
+    },
+    {
+      "epoch": 0.2241207145234661,
+      "grad_norm": 1.71875,
+      "learning_rate": 2.9932998216189077e-05,
+      "loss": 0.423,
+      "step": 2020
+    },
+    {
+      "epoch": 0.22633973149894596,
+      "grad_norm": 1.703125,
+      "learning_rate": 2.992008998928534e-05,
+      "loss": 0.4803,
+      "step": 2040
+    },
+    {
+      "epoch": 0.22855874847442584,
+      "grad_norm": 1.4296875,
+      "learning_rate": 2.990604868935424e-05,
+      "loss": 0.4604,
+      "step": 2060
+    },
+    {
+      "epoch": 0.2307777654499057,
+      "grad_norm": 1.890625,
+      "learning_rate": 2.989087538273105e-05,
+      "loss": 0.5155,
+      "step": 2080
+    },
+    {
+      "epoch": 0.23299678242538555,
+      "grad_norm": 1.7421875,
+      "learning_rate": 2.9874571221718776e-05,
+      "loss": 0.5122,
+      "step": 2100
+    },
+    {
+      "epoch": 0.23521579940086543,
+      "grad_norm": 1.765625,
+      "learning_rate": 2.985713744450063e-05,
+      "loss": 0.4475,
+      "step": 2120
+    },
+    {
+      "epoch": 0.23743481637634528,
+      "grad_norm": 2.15625,
+      "learning_rate": 2.9838575375046034e-05,
+      "loss": 0.4281,
+      "step": 2140
+    },
+    {
+      "epoch": 0.23965383335182514,
+      "grad_norm": 1.9296875,
+      "learning_rate": 2.9818886423010024e-05,
+      "loss": 0.4383,
+      "step": 2160
+    },
+    {
+      "epoch": 0.24187285032730502,
+      "grad_norm": 1.7421875,
+      "learning_rate": 2.979807208362625e-05,
+      "loss": 0.449,
+      "step": 2180
+    },
+    {
+      "epoch": 0.24409186730278487,
+      "grad_norm": 1.4453125,
+      "learning_rate": 2.9776133937593395e-05,
+      "loss": 0.5002,
+      "step": 2200
+    },
+    {
+      "epoch": 0.24631088427826472,
+      "grad_norm": 1.6484375,
+      "learning_rate": 2.9753073650955128e-05,
+      "loss": 0.4667,
+      "step": 2220
+    },
+    {
+      "epoch": 0.24852990125374458,
+      "grad_norm": 2.203125,
+      "learning_rate": 2.9728892974973592e-05,
+      "loss": 0.4431,
+      "step": 2240
+    },
+    {
+      "epoch": 0.25074891822922446,
+      "grad_norm": 1.78125,
+      "learning_rate": 2.970359374599641e-05,
+      "loss": 0.4684,
+      "step": 2260
+    },
+    {
+      "epoch": 0.25296793520470434,
+      "grad_norm": 2.15625,
+      "learning_rate": 2.967717788531722e-05,
+      "loss": 0.4599,
+      "step": 2280
+    },
+    {
+      "epoch": 0.25518695218018417,
+      "grad_norm": 1.4453125,
+      "learning_rate": 2.9649647399029764e-05,
+      "loss": 0.5047,
+      "step": 2300
+    },
+    {
+      "epoch": 0.25740596915566405,
+      "grad_norm": 2.265625,
+      "learning_rate": 2.9621004377875558e-05,
+      "loss": 0.4824,
+      "step": 2320
+    },
+    {
+      "epoch": 0.2596249861311439,
+      "grad_norm": 1.3515625,
+      "learning_rate": 2.959125099708509e-05,
+      "loss": 0.4928,
+      "step": 2340
+    },
+    {
+      "epoch": 0.26184400310662376,
+      "grad_norm": 1.5703125,
+      "learning_rate": 2.9560389516212638e-05,
+      "loss": 0.5386,
+      "step": 2360
+    },
+    {
+      "epoch": 0.26406302008210364,
+      "grad_norm": 0.96875,
+      "learning_rate": 2.9528422278964687e-05,
+      "loss": 0.4092,
+      "step": 2380
+    },
+    {
+      "epoch": 0.26628203705758346,
+      "grad_norm": 1.5703125,
+      "learning_rate": 2.949535171302192e-05,
+      "loss": 0.4929,
+      "step": 2400
+    },
+    {
+      "epoch": 0.26850105403306335,
+      "grad_norm": 1.5859375,
+      "learning_rate": 2.9461180329854867e-05,
+      "loss": 0.4893,
+      "step": 2420
+    },
+    {
+      "epoch": 0.27072007100854323,
+      "grad_norm": 1.890625,
+      "learning_rate": 2.9425910724533165e-05,
+      "loss": 0.4554,
+      "step": 2440
+    },
+    {
+      "epoch": 0.27293908798402305,
+      "grad_norm": 1.65625,
+      "learning_rate": 2.9389545575528496e-05,
+      "loss": 0.4702,
+      "step": 2460
+    },
+    {
+      "epoch": 0.27515810495950294,
+      "grad_norm": 1.375,
+      "learning_rate": 2.9352087644511162e-05,
+      "loss": 0.5056,
+      "step": 2480
+    },
+    {
+      "epoch": 0.2773771219349828,
+      "grad_norm": 2.859375,
+      "learning_rate": 2.9313539776140362e-05,
+      "loss": 0.428,
+      "step": 2500
+    },
+    {
+      "epoch": 0.27959613891046264,
+      "grad_norm": 1.7890625,
+      "learning_rate": 2.9273904897848174e-05,
+      "loss": 0.4827,
+      "step": 2520
+    },
+    {
+      "epoch": 0.2818151558859425,
+      "grad_norm": 1.359375,
+      "learning_rate": 2.9233186019617214e-05,
+      "loss": 0.5082,
+      "step": 2540
+    },
+    {
+      "epoch": 0.2840341728614224,
+      "grad_norm": 1.421875,
+      "learning_rate": 2.9191386233752062e-05,
+      "loss": 0.4803,
+      "step": 2560
+    },
+    {
+      "epoch": 0.28625318983690223,
+      "grad_norm": 1.859375,
+      "learning_rate": 2.9148508714644424e-05,
+      "loss": 0.5205,
+      "step": 2580
+    },
+    {
+      "epoch": 0.2884722068123821,
+      "grad_norm": 1.546875,
+      "learning_rate": 2.9104556718532054e-05,
+      "loss": 0.4965,
+      "step": 2600
+    },
+    {
+      "epoch": 0.290691223787862,
+      "grad_norm": 1.5078125,
+      "learning_rate": 2.9059533583251487e-05,
+      "loss": 0.4376,
+      "step": 2620
+    },
+    {
+      "epoch": 0.2929102407633418,
+      "grad_norm": 1.7578125,
+      "learning_rate": 2.9013442727984517e-05,
+      "loss": 0.4618,
+      "step": 2640
+    },
+    {
+      "epoch": 0.2951292577388217,
+      "grad_norm": 1.640625,
+      "learning_rate": 2.8966287652998562e-05,
+      "loss": 0.4943,
+      "step": 2660
+    },
+    {
+      "epoch": 0.2973482747143016,
+      "grad_norm": 1.703125,
+      "learning_rate": 2.891807193938085e-05,
+      "loss": 0.4582,
+      "step": 2680
+    },
+    {
+      "epoch": 0.2995672916897814,
+      "grad_norm": 1.6875,
+      "learning_rate": 2.8868799248766436e-05,
+      "loss": 0.5133,
+      "step": 2700
+    },
+    {
+      "epoch": 0.3017863086652613,
+      "grad_norm": 1.71875,
+      "learning_rate": 2.8818473323060143e-05,
+      "loss": 0.4628,
+      "step": 2720
+    },
+    {
+      "epoch": 0.3040053256407412,
+      "grad_norm": 1.609375,
+      "learning_rate": 2.87670979841524e-05,
+      "loss": 0.4408,
+      "step": 2740
+    },
+    {
+      "epoch": 0.306224342616221,
+      "grad_norm": 1.734375,
+      "learning_rate": 2.8714677133628963e-05,
+      "loss": 0.5004,
+      "step": 2760
+    },
+    {
+      "epoch": 0.3084433595917009,
+      "grad_norm": 1.6640625,
+      "learning_rate": 2.866121475247467e-05,
+      "loss": 0.4031,
+      "step": 2780
+    },
+    {
+      "epoch": 0.31066237656718076,
+      "grad_norm": 1.8046875,
+      "learning_rate": 2.8606714900771055e-05,
+      "loss": 0.5192,
+      "step": 2800
+    },
+    {
+      "epoch": 0.3128813935426606,
+      "grad_norm": 2.03125,
+      "learning_rate": 2.8551181717388066e-05,
+      "loss": 0.5397,
+      "step": 2820
+    },
+    {
+      "epoch": 0.3151004105181405,
+      "grad_norm": 1.9609375,
+      "learning_rate": 2.849461941966972e-05,
+      "loss": 0.4446,
+      "step": 2840
+    },
+    {
+      "epoch": 0.3173194274936203,
+      "grad_norm": 1.9296875,
+      "learning_rate": 2.8437032303113823e-05,
+      "loss": 0.4464,
+      "step": 2860
+    },
+    {
+      "epoch": 0.3195384444691002,
+      "grad_norm": 1.734375,
+      "learning_rate": 2.8378424741045773e-05,
+      "loss": 0.526,
+      "step": 2880
+    },
+    {
+      "epoch": 0.32175746144458006,
+      "grad_norm": 1.7265625,
+      "learning_rate": 2.831880118428644e-05,
+      "loss": 0.4087,
+      "step": 2900
+    },
+    {
+      "epoch": 0.3239764784200599,
+      "grad_norm": 1.125,
+      "learning_rate": 2.8258166160814135e-05,
+      "loss": 0.4833,
+      "step": 2920
+    },
+    {
+      "epoch": 0.32619549539553977,
+      "grad_norm": 1.34375,
+      "learning_rate": 2.8196524275420758e-05,
+      "loss": 0.4403,
+      "step": 2940
+    },
+    {
+      "epoch": 0.32841451237101965,
+      "grad_norm": 1.7421875,
+      "learning_rate": 2.813388020936211e-05,
+      "loss": 0.4728,
+      "step": 2960
+    },
+    {
+      "epoch": 0.3306335293464995,
+      "grad_norm": 1.0390625,
+      "learning_rate": 2.8070238720002364e-05,
+      "loss": 0.4389,
+      "step": 2980
+    },
+    {
+      "epoch": 0.33285254632197936,
+      "grad_norm": 0.90625,
+      "learning_rate": 2.800560464045278e-05,
+      "loss": 0.482,
+      "step": 3000
+    },
+    {
+      "epoch": 0.33507156329745924,
+      "grad_norm": 1.6796875,
+      "learning_rate": 2.7939982879204686e-05,
+      "loss": 0.4157,
+      "step": 3020
+    },
+    {
+      "epoch": 0.33729058027293907,
+      "grad_norm": 1.9140625,
+      "learning_rate": 2.787337841975668e-05,
+      "loss": 0.4593,
+      "step": 3040
+    },
+    {
+      "epoch": 0.33950959724841895,
+      "grad_norm": 1.2421875,
+      "learning_rate": 2.78057963202362e-05,
+      "loss": 0.4071,
+      "step": 3060
+    },
+    {
+      "epoch": 0.34172861422389883,
+      "grad_norm": 1.671875,
+      "learning_rate": 2.773724171301538e-05,
+      "loss": 0.4307,
+      "step": 3080
+    },
+    {
+      "epoch": 0.34394763119937866,
+      "grad_norm": 2.75,
+      "learning_rate": 2.7667719804321285e-05,
+      "loss": 0.5196,
+      "step": 3100
+    },
+    {
+      "epoch": 0.34616664817485854,
+      "grad_norm": 1.71875,
+      "learning_rate": 2.7597235873840544e-05,
+      "loss": 0.4529,
+      "step": 3120
+    },
+    {
+      "epoch": 0.3483856651503384,
+      "grad_norm": 1.453125,
+      "learning_rate": 2.7525795274318386e-05,
+      "loss": 0.5671,
+      "step": 3140
+    },
+    {
+      "epoch": 0.35060468212581825,
+      "grad_norm": 1.6484375,
+      "learning_rate": 2.745340343115213e-05,
+      "loss": 0.506,
+      "step": 3160
+    },
+    {
+      "epoch": 0.3528236991012981,
+      "grad_norm": 1.828125,
+      "learning_rate": 2.7380065841979196e-05,
+      "loss": 0.4834,
+      "step": 3180
+    },
+    {
+      "epoch": 0.355042716076778,
+      "grad_norm": 1.9296875,
+      "learning_rate": 2.7305788076259565e-05,
+      "loss": 0.4635,
+      "step": 3200
+    },
+    {
+      "epoch": 0.35726173305225783,
+      "grad_norm": 1.7109375,
+      "learning_rate": 2.7230575774852843e-05,
+      "loss": 0.4873,
+      "step": 3220
+    },
+    {
+      "epoch": 0.3594807500277377,
+      "grad_norm": 1.578125,
+      "learning_rate": 2.715443464958986e-05,
+      "loss": 0.4133,
+      "step": 3240
+    },
+    {
+      "epoch": 0.3616997670032176,
+      "grad_norm": 2.03125,
+      "learning_rate": 2.707737048283891e-05,
+      "loss": 0.4954,
+      "step": 3260
+    },
+    {
+      "epoch": 0.3639187839786974,
+      "grad_norm": 1.3671875,
+      "learning_rate": 2.699938912706663e-05,
+      "loss": 0.4585,
+      "step": 3280
+    },
+    {
+      "epoch": 0.3661378009541773,
+      "grad_norm": 1.921875,
+      "learning_rate": 2.6920496504393507e-05,
+      "loss": 0.5479,
+      "step": 3300
+    },
+    {
+      "epoch": 0.3683568179296572,
+      "grad_norm": 2.515625,
+      "learning_rate": 2.6840698606144197e-05,
+      "loss": 0.4712,
+      "step": 3320
+    },
+    {
+      "epoch": 0.370575834905137,
+      "grad_norm": 2.28125,
+      "learning_rate": 2.6760001492392474e-05,
+      "loss": 0.5046,
+      "step": 3340
+    },
+    {
+      "epoch": 0.3727948518806169,
+      "grad_norm": 1.53125,
+      "learning_rate": 2.6678411291501038e-05,
+      "loss": 0.522,
+      "step": 3360
+    },
+    {
+      "epoch": 0.3750138688560967,
+      "grad_norm": 1.6796875,
+      "learning_rate": 2.6595934199656108e-05,
+      "loss": 0.4852,
+      "step": 3380
+    },
+    {
+      "epoch": 0.3772328858315766,
+      "grad_norm": 1.9140625,
+      "learning_rate": 2.6512576480396862e-05,
+      "loss": 0.4745,
+      "step": 3400
+    },
+    {
+      "epoch": 0.3794519028070565,
+      "grad_norm": 1.71875,
+      "learning_rate": 2.6428344464139756e-05,
+      "loss": 0.4903,
+      "step": 3420
+    },
+    {
+      "epoch": 0.3816709197825363,
+      "grad_norm": 1.625,
+      "learning_rate": 2.6343244547697798e-05,
+      "loss": 0.3906,
+      "step": 3440
+    },
+    {
+      "epoch": 0.3838899367580162,
+      "grad_norm": 1.859375,
+      "learning_rate": 2.6257283193794742e-05,
+      "loss": 0.471,
+      "step": 3460
+    },
+    {
+      "epoch": 0.3861089537334961,
+      "grad_norm": 1.90625,
+      "learning_rate": 2.617046693057429e-05,
+      "loss": 0.4878,
+      "step": 3480
+    },
+    {
+      "epoch": 0.3883279707089759,
+      "grad_norm": 1.765625,
+      "learning_rate": 2.6082802351104317e-05,
+      "loss": 0.4718,
+      "step": 3500
+    },
+    {
+      "epoch": 0.3905469876844558,
+      "grad_norm": 1.8828125,
+      "learning_rate": 2.5994296112876222e-05,
+      "loss": 0.4649,
+      "step": 3520
+    },
+    {
+      "epoch": 0.39276600465993566,
+      "grad_norm": 1.2109375,
+      "learning_rate": 2.5904954937299267e-05,
+      "loss": 0.3973,
+      "step": 3540
+    },
+    {
+      "epoch": 0.3949850216354155,
+      "grad_norm": 1.4609375,
+      "learning_rate": 2.5814785609190197e-05,
+      "loss": 0.4785,
+      "step": 3560
+    },
+    {
+      "epoch": 0.39720403861089537,
+      "grad_norm": 2.171875,
+      "learning_rate": 2.5723794976257947e-05,
+      "loss": 0.4512,
+      "step": 3580
+    },
+    {
+      "epoch": 0.39942305558637525,
+      "grad_norm": 1.7265625,
+      "learning_rate": 2.5631989948583623e-05,
+      "loss": 0.4931,
+      "step": 3600
+    },
+    {
+      "epoch": 0.4016420725618551,
+      "grad_norm": 1.1796875,
+      "learning_rate": 2.553937749809572e-05,
+      "loss": 0.4183,
+      "step": 3620
+    },
+    {
+      "epoch": 0.40386108953733496,
+      "grad_norm": 1.7890625,
+      "learning_rate": 2.544596465804068e-05,
+      "loss": 0.4321,
+      "step": 3640
+    },
+    {
+      "epoch": 0.40608010651281484,
+      "grad_norm": 2.109375,
+      "learning_rate": 2.5351758522448724e-05,
+      "loss": 0.489,
+      "step": 3660
+    },
+    {
+      "epoch": 0.40829912348829467,
+      "grad_norm": 2.03125,
+      "learning_rate": 2.5256766245595166e-05,
+      "loss": 0.4398,
+      "step": 3680
+    },
+    {
+      "epoch": 0.41051814046377455,
+      "grad_norm": 1.6484375,
+      "learning_rate": 2.516099504145703e-05,
+      "loss": 0.4831,
+      "step": 3700
+    },
+    {
+      "epoch": 0.41273715743925443,
+      "grad_norm": 1.90625,
+      "learning_rate": 2.5064452183165283e-05,
+      "loss": 0.4365,
+      "step": 3720
+    },
+    {
+      "epoch": 0.41495617441473426,
+      "grad_norm": 1.359375,
+      "learning_rate": 2.496714500245241e-05,
+      "loss": 0.4309,
+      "step": 3740
+    },
+    {
+      "epoch": 0.41717519139021414,
+      "grad_norm": 1.765625,
+      "learning_rate": 2.4869080889095693e-05,
+      "loss": 0.5378,
+      "step": 3760
+    },
+    {
+      "epoch": 0.419394208365694,
+      "grad_norm": 1.75,
+      "learning_rate": 2.477026729035595e-05,
+      "loss": 0.4905,
+      "step": 3780
+    },
+    {
+      "epoch": 0.42161322534117385,
+      "grad_norm": 1.859375,
+      "learning_rate": 2.4670711710412026e-05,
+      "loss": 0.4187,
+      "step": 3800
+    },
+    {
+      "epoch": 0.42383224231665373,
+      "grad_norm": 1.9140625,
+      "learning_rate": 2.457042170979086e-05,
+      "loss": 0.4817,
+      "step": 3820
+    },
+    {
+      "epoch": 0.4260512592921336,
+      "grad_norm": 1.6953125,
+      "learning_rate": 2.4469404904793338e-05,
+      "loss": 0.5108,
+      "step": 3840
+    },
+    {
+      "epoch": 0.42827027626761344,
+      "grad_norm": 1.46875,
+      "learning_rate": 2.4367668966915885e-05,
+      "loss": 0.5112,
+      "step": 3860
+    },
+    {
+      "epoch": 0.4304892932430933,
+      "grad_norm": 3.046875,
+      "learning_rate": 2.4265221622267876e-05,
+      "loss": 0.5353,
+      "step": 3880
+    },
+    {
+      "epoch": 0.43270831021857314,
+      "grad_norm": 1.6640625,
+      "learning_rate": 2.4162070650984893e-05,
+      "loss": 0.4684,
+      "step": 3900
+    },
+    {
+      "epoch": 0.434927327194053,
+      "grad_norm": 1.5078125,
+      "learning_rate": 2.4058223886637872e-05,
+      "loss": 0.4374,
+      "step": 3920
+    },
+    {
+      "epoch": 0.4371463441695329,
+      "grad_norm": 1.6953125,
+      "learning_rate": 2.3953689215638194e-05,
+      "loss": 0.3753,
+      "step": 3940
+    },
+    {
+      "epoch": 0.43936536114501273,
+      "grad_norm": 1.4921875,
+      "learning_rate": 2.3848474576638807e-05,
+      "loss": 0.4427,
+      "step": 3960
+    },
+    {
+      "epoch": 0.4415843781204926,
+      "grad_norm": 1.25,
+      "learning_rate": 2.3742587959931285e-05,
+      "loss": 0.5074,
+      "step": 3980
+    },
+    {
+      "epoch": 0.4438033950959725,
+      "grad_norm": 1.7265625,
+      "learning_rate": 2.3636037406839076e-05,
+      "loss": 0.4841,
+      "step": 4000
+    },
+    {
+      "epoch": 0.4460224120714523,
+      "grad_norm": 1.8203125,
+      "learning_rate": 2.3528831009106786e-05,
+      "loss": 0.4643,
+      "step": 4020
+    },
+    {
+      "epoch": 0.4482414290469322,
+      "grad_norm": 1.25,
+      "learning_rate": 2.3420976908285687e-05,
+      "loss": 0.4976,
+      "step": 4040
+    },
+    {
+      "epoch": 0.4504604460224121,
+      "grad_norm": 1.828125,
+      "learning_rate": 2.3312483295115424e-05,
+      "loss": 0.5453,
+      "step": 4060
+    },
+    {
+      "epoch": 0.4526794629978919,
+      "grad_norm": 2.921875,
+      "learning_rate": 2.320335840890198e-05,
+      "loss": 0.46,
+      "step": 4080
+    },
+    {
+      "epoch": 0.4548984799733718,
+      "grad_norm": 2.03125,
+      "learning_rate": 2.3093610536891965e-05,
+      "loss": 0.5129,
+      "step": 4100
+    },
+    {
+      "epoch": 0.4571174969488517,
+      "grad_norm": 1.4375,
+      "learning_rate": 2.2983248013643253e-05,
+      "loss": 0.4429,
+      "step": 4120
+    },
+    {
+      "epoch": 0.4593365139243315,
+      "grad_norm": 2.03125,
+      "learning_rate": 2.2872279220392054e-05,
+      "loss": 0.4817,
+      "step": 4140
+    },
+    {
+      "epoch": 0.4615555308998114,
+      "grad_norm": 1.8984375,
+      "learning_rate": 2.2760712584416386e-05,
+      "loss": 0.51,
+      "step": 4160
+    },
+    {
+      "epoch": 0.46377454787529127,
+      "grad_norm": 1.2734375,
+      "learning_rate": 2.2648556578396107e-05,
+      "loss": 0.5001,
+      "step": 4180
+    },
+    {
+      "epoch": 0.4659935648507711,
+      "grad_norm": 2.34375,
+      "learning_rate": 2.2535819719769487e-05,
+      "loss": 0.4739,
+      "step": 4200
+    },
+    {
+      "epoch": 0.468212581826251,
+      "grad_norm": 1.484375,
+      "learning_rate": 2.242251057008633e-05,
+      "loss": 0.474,
+      "step": 4220
+    },
+    {
+      "epoch": 0.47043159880173085,
+      "grad_norm": 1.7421875,
+      "learning_rate": 2.2308637734357826e-05,
+      "loss": 0.4622,
+      "step": 4240
+    },
+    {
+      "epoch": 0.4726506157772107,
+      "grad_norm": 1.4296875,
+      "learning_rate": 2.219420986040305e-05,
+      "loss": 0.5075,
+      "step": 4260
+    },
+    {
+      "epoch": 0.47486963275269056,
+      "grad_norm": 1.7734375,
+      "learning_rate": 2.2079235638192203e-05,
+      "loss": 0.4817,
+      "step": 4280
+    },
+    {
+      "epoch": 0.47708864972817044,
+      "grad_norm": 1.3671875,
+      "learning_rate": 2.1963723799186706e-05,
+      "loss": 0.445,
+      "step": 4300
+    },
+    {
+      "epoch": 0.47930766670365027,
+      "grad_norm": 1.6796875,
+      "learning_rate": 2.184768311567608e-05,
+      "loss": 0.475,
+      "step": 4320
+    },
+    {
+      "epoch": 0.48152668367913015,
+      "grad_norm": 1.640625,
+      "learning_rate": 2.1731122400111764e-05,
+      "loss": 0.4418,
+      "step": 4340
+    },
+    {
+      "epoch": 0.48374570065461003,
+      "grad_norm": 1.765625,
+      "learning_rate": 2.161405050443789e-05,
+      "loss": 0.4722,
+      "step": 4360
+    },
+    {
+      "epoch": 0.48596471763008986,
+      "grad_norm": 1.75,
+      "learning_rate": 2.1496476319419002e-05,
+      "loss": 0.4814,
+      "step": 4380
+    },
+    {
+      "epoch": 0.48818373460556974,
+      "grad_norm": 1.59375,
+      "learning_rate": 2.137840877396491e-05,
+      "loss": 0.5435,
+      "step": 4400
+    },
+    {
+      "epoch": 0.49040275158104957,
+      "grad_norm": 1.734375,
+      "learning_rate": 2.125985683445258e-05,
+      "loss": 0.4521,
+      "step": 4420
+    },
+    {
+      "epoch": 0.49262176855652945,
+      "grad_norm": 1.6640625,
+      "learning_rate": 2.114082950404519e-05,
+      "loss": 0.452,
+      "step": 4440
+    },
+    {
+      "epoch": 0.49484078553200933,
+      "grad_norm": 1.2734375,
+      "learning_rate": 2.1021335822008447e-05,
+      "loss": 0.5176,
+      "step": 4460
+    },
+    {
+      "epoch": 0.49705980250748916,
+      "grad_norm": 1.6328125,
+      "learning_rate": 2.0901384863024078e-05,
+      "loss": 0.4307,
+      "step": 4480
+    },
+    {
+      "epoch": 0.49927881948296904,
+      "grad_norm": 1.6640625,
+      "learning_rate": 2.0780985736500696e-05,
+      "loss": 0.4856,
+      "step": 4500
+    },
+    {
+      "epoch": 0.5014978364584489,
+      "grad_norm": 1.9609375,
+      "learning_rate": 2.0660147585881994e-05,
+      "loss": 0.4339,
+      "step": 4520
+    },
+    {
+      "epoch": 0.5037168534339288,
+      "grad_norm": 2.046875,
+      "learning_rate": 2.0538879587952382e-05,
+      "loss": 0.4902,
+      "step": 4540
+    },
+    {
+      "epoch": 0.5059358704094087,
+      "grad_norm": 1.8671875,
+      "learning_rate": 2.0417190952140064e-05,
+      "loss": 0.4343,
+      "step": 4560
+    },
+    {
+      "epoch": 0.5081548873848885,
+      "grad_norm": 1.7265625,
+      "learning_rate": 2.029509091981765e-05,
+      "loss": 0.4416,
+      "step": 4580
+    },
+    {
+      "epoch": 0.5103739043603683,
+      "grad_norm": 1.9375,
+      "learning_rate": 2.0172588763600335e-05,
+      "loss": 0.4802,
+      "step": 4600
+    },
+    {
+      "epoch": 0.5125929213358482,
+      "grad_norm": 1.703125,
+      "learning_rate": 2.0049693786641734e-05,
+      "loss": 0.4794,
+      "step": 4620
+    },
+    {
+      "epoch": 0.5148119383113281,
+      "grad_norm": 2.8125,
+      "learning_rate": 1.9926415321927347e-05,
+      "loss": 0.5092,
+      "step": 4640
+    },
+    {
+      "epoch": 0.517030955286808,
+      "grad_norm": 1.3984375,
+      "learning_rate": 1.980276273156581e-05,
+      "loss": 0.4595,
+      "step": 4660
+    },
+    {
+      "epoch": 0.5192499722622878,
+      "grad_norm": 1.484375,
+      "learning_rate": 1.9678745406077886e-05,
+      "loss": 0.452,
+      "step": 4680
+    },
+    {
+      "epoch": 0.5214689892377676,
+      "grad_norm": 1.703125,
+      "learning_rate": 1.9554372763683337e-05,
+      "loss": 0.4528,
+      "step": 4700
+    },
+    {
+      "epoch": 0.5236880062132475,
+      "grad_norm": 1.5,
+      "learning_rate": 1.9429654249585684e-05,
+      "loss": 0.4743,
+      "step": 4720
+    },
+    {
+      "epoch": 0.5259070231887274,
+      "grad_norm": 1.5703125,
+      "learning_rate": 1.9304599335254894e-05,
+      "loss": 0.4476,
+      "step": 4740
+    },
+    {
+      "epoch": 0.5281260401642073,
+      "grad_norm": 1.546875,
+      "learning_rate": 1.9179217517708117e-05,
+      "loss": 0.447,
+      "step": 4760
+    },
+    {
+      "epoch": 0.5303450571396872,
+      "grad_norm": 1.625,
+      "learning_rate": 1.9053518318788428e-05,
+      "loss": 0.4606,
+      "step": 4780
+    },
+    {
+      "epoch": 0.5325640741151669,
+      "grad_norm": 1.6875,
+      "learning_rate": 1.8927511284441722e-05,
+      "loss": 0.4561,
+      "step": 4800
+    },
+    {
+      "epoch": 0.5347830910906468,
+      "grad_norm": 1.8359375,
+      "learning_rate": 1.880120598399178e-05,
+      "loss": 0.4522,
+      "step": 4820
+    },
+    {
+      "epoch": 0.5370021080661267,
+      "grad_norm": 1.9921875,
+      "learning_rate": 1.8674612009413536e-05,
+      "loss": 0.4943,
+      "step": 4840
+    },
+    {
+      "epoch": 0.5392211250416066,
+      "grad_norm": 1.703125,
+      "learning_rate": 1.8547738974604623e-05,
+      "loss": 0.5822,
+      "step": 4860
+    },
+    {
+      "epoch": 0.5414401420170865,
+      "grad_norm": 1.7734375,
+      "learning_rate": 1.842059651465531e-05,
+      "loss": 0.4273,
+      "step": 4880
+    },
+    {
+      "epoch": 0.5436591589925663,
+      "grad_norm": 1.6953125,
+      "learning_rate": 1.829319428511673e-05,
+      "loss": 0.4704,
+      "step": 4900
+    },
+    {
+      "epoch": 0.5458781759680461,
+      "grad_norm": 1.671875,
+      "learning_rate": 1.816554196126767e-05,
+      "loss": 0.5452,
+      "step": 4920
+    },
+    {
+      "epoch": 0.548097192943526,
+      "grad_norm": 1.8203125,
+      "learning_rate": 1.803764923737974e-05,
+      "loss": 0.436,
+      "step": 4940
+    },
+    {
+      "epoch": 0.5503162099190059,
+      "grad_norm": 1.7578125,
+      "learning_rate": 1.7909525825981214e-05,
+      "loss": 0.4518,
+      "step": 4960
+    },
+    {
+      "epoch": 0.5525352268944858,
+      "grad_norm": 2.03125,
+      "learning_rate": 1.778118145711942e-05,
+      "loss": 0.4138,
+      "step": 4980
+    },
+    {
+      "epoch": 0.5547542438699656,
+      "grad_norm": 2.46875,
+      "learning_rate": 1.7652625877621793e-05,
+      "loss": 0.4332,
+      "step": 5000
+    },
+    {
+      "epoch": 0.5569732608454455,
+      "grad_norm": 2.09375,
+      "learning_rate": 1.7523868850355704e-05,
+      "loss": 0.4842,
+      "step": 5020
+    },
+    {
+      "epoch": 0.5591922778209253,
+      "grad_norm": 1.6328125,
+      "learning_rate": 1.7394920153487022e-05,
+      "loss": 0.4935,
+      "step": 5040
+    },
+    {
+      "epoch": 0.5614112947964052,
+      "grad_norm": 1.6953125,
+      "learning_rate": 1.7265789579737528e-05,
+      "loss": 0.5129,
+      "step": 5060
+    },
+    {
+      "epoch": 0.563630311771885,
+      "grad_norm": 1.65625,
+      "learning_rate": 1.7136486935641256e-05,
+      "loss": 0.4281,
+      "step": 5080
+    },
+    {
+      "epoch": 0.5658493287473649,
+      "grad_norm": 2.015625,
+      "learning_rate": 1.7007022040799726e-05,
+      "loss": 0.4634,
+      "step": 5100
+    },
+    {
+      "epoch": 0.5680683457228448,
+      "grad_norm": 2.15625,
+      "learning_rate": 1.687740472713623e-05,
+      "loss": 0.5225,
+      "step": 5120
+    },
+    {
+      "epoch": 0.5702873626983247,
+      "grad_norm": 1.9921875,
+      "learning_rate": 1.674764483814918e-05,
+      "loss": 0.4838,
+      "step": 5140
+    },
+    {
+      "epoch": 0.5725063796738045,
+      "grad_norm": 1.7421875,
+      "learning_rate": 1.661775222816453e-05,
+      "loss": 0.4291,
+      "step": 5160
+    },
+    {
+      "epoch": 0.5747253966492843,
+      "grad_norm": 1.671875,
+      "learning_rate": 1.648773676158747e-05,
+      "loss": 0.3925,
+      "step": 5180
+    },
+    {
+      "epoch": 0.5769444136247642,
+      "grad_norm": 1.734375,
+      "learning_rate": 1.6357608312153223e-05,
+      "loss": 0.4385,
+      "step": 5200
+    },
+    {
+      "epoch": 0.5791634306002441,
+      "grad_norm": 1.71875,
+      "learning_rate": 1.6227376762177272e-05,
+      "loss": 0.4144,
+      "step": 5220
+    },
+    {
+      "epoch": 0.581382447575724,
+      "grad_norm": 2.71875,
+      "learning_rate": 1.6097052001804825e-05,
+      "loss": 0.5116,
+      "step": 5240
+    },
+    {
+      "epoch": 0.5836014645512038,
+      "grad_norm": 1.953125,
+      "learning_rate": 1.5966643928259753e-05,
+      "loss": 0.4869,
+      "step": 5260
+    },
+    {
+      "epoch": 0.5858204815266836,
+      "grad_norm": 1.9140625,
+      "learning_rate": 1.5836162445092963e-05,
+      "loss": 0.4644,
+      "step": 5280
+    },
+    {
+      "epoch": 0.5880394985021635,
+      "grad_norm": 1.8203125,
+      "learning_rate": 1.5705617461430282e-05,
+      "loss": 0.3844,
+      "step": 5300
+    },
+    {
+      "epoch": 0.5902585154776434,
+      "grad_norm": 1.75,
+      "learning_rate": 1.5575018891219944e-05,
+      "loss": 0.4233,
+      "step": 5320
+    },
+    {
+      "epoch": 0.5924775324531233,
+      "grad_norm": 1.140625,
+      "learning_rate": 1.5444376652479706e-05,
+      "loss": 0.5273,
+      "step": 5340
+    },
+    {
+      "epoch": 0.5946965494286032,
+      "grad_norm": 1.921875,
+      "learning_rate": 1.531370066654362e-05,
+      "loss": 0.4783,
+      "step": 5360
+    },
+    {
+      "epoch": 0.5969155664040829,
+      "grad_norm": 1.796875,
+      "learning_rate": 1.5183000857308604e-05,
+      "loss": 0.4747,
+      "step": 5380
+    },
+    {
+      "epoch": 0.5991345833795628,
+      "grad_norm": 1.6484375,
+      "learning_rate": 1.5052287150480774e-05,
+      "loss": 0.4335,
+      "step": 5400
+    },
+    {
+      "epoch": 0.6013536003550427,
+      "grad_norm": 1.75,
+      "learning_rate": 1.4921569472821673e-05,
+      "loss": 0.4489,
+      "step": 5420
+    },
+    {
+      "epoch": 0.6035726173305226,
+      "grad_norm": 2.21875,
+      "learning_rate": 1.4790857751394398e-05,
+      "loss": 0.4495,
+      "step": 5440
+    },
+    {
+      "epoch": 0.6057916343060025,
+      "grad_norm": 1.5703125,
+      "learning_rate": 1.4660161912809718e-05,
+      "loss": 0.4516,
+      "step": 5460
+    },
+    {
+      "epoch": 0.6080106512814824,
+      "grad_norm": 1.4765625,
+      "learning_rate": 1.4529491882472209e-05,
+      "loss": 0.4418,
+      "step": 5480
+    },
+    {
+      "epoch": 0.6102296682569621,
+      "grad_norm": 1.6015625,
+      "learning_rate": 1.4398857583826501e-05,
+      "loss": 0.4701,
+      "step": 5500
+    },
+    {
+      "epoch": 0.612448685232442,
+      "grad_norm": 2.390625,
+      "learning_rate": 1.4268268937603659e-05,
+      "loss": 0.4957,
+      "step": 5520
+    },
+    {
+      "epoch": 0.6146677022079219,
+      "grad_norm": 1.7421875,
+      "learning_rate": 1.413773586106777e-05,
+      "loss": 0.5176,
+      "step": 5540
+    },
+    {
+      "epoch": 0.6168867191834018,
+      "grad_norm": 1.8125,
+      "learning_rate": 1.400726826726282e-05,
+      "loss": 0.4252,
+      "step": 5560
+    },
+    {
+      "epoch": 0.6191057361588816,
+      "grad_norm": 1.4453125,
+      "learning_rate": 1.3876876064259836e-05,
+      "loss": 0.473,
+      "step": 5580
+    },
+    {
+      "epoch": 0.6213247531343615,
+      "grad_norm": 1.6796875,
+      "learning_rate": 1.3746569154404477e-05,
+      "loss": 0.4589,
+      "step": 5600
+    },
+    {
+      "epoch": 0.6235437701098413,
+      "grad_norm": 2.734375,
+      "learning_rate": 1.3616357433564993e-05,
+      "loss": 0.4083,
+      "step": 5620
+    },
+    {
+      "epoch": 0.6257627870853212,
+      "grad_norm": 1.75,
+      "learning_rate": 1.348625079038071e-05,
+      "loss": 0.4638,
+      "step": 5640
+    },
+    {
+      "epoch": 0.6279818040608011,
+      "grad_norm": 2.0625,
+      "learning_rate": 1.335625910551108e-05,
+      "loss": 0.4703,
+      "step": 5660
+    },
+    {
+      "epoch": 0.630200821036281,
+      "grad_norm": 1.7109375,
+      "learning_rate": 1.3226392250885288e-05,
+      "loss": 0.4683,
+      "step": 5680
+    },
+    {
+      "epoch": 0.6324198380117608,
+      "grad_norm": 1.2265625,
+      "learning_rate": 1.3096660088952581e-05,
+      "loss": 0.4641,
+      "step": 5700
+    },
+    {
+      "epoch": 0.6346388549872406,
+      "grad_norm": 1.6953125,
+      "learning_rate": 1.2967072471933255e-05,
+      "loss": 0.4324,
+      "step": 5720
+    },
+    {
+      "epoch": 0.6368578719627205,
+      "grad_norm": 1.6953125,
+      "learning_rate": 1.283763924107046e-05,
+      "loss": 0.5183,
+      "step": 5740
+    },
+    {
+      "epoch": 0.6390768889382004,
+      "grad_norm": 1.4921875,
+      "learning_rate": 1.2708370225882848e-05,
+      "loss": 0.4178,
+      "step": 5760
+    },
+    {
+      "epoch": 0.6412959059136802,
+      "grad_norm": 2.046875,
+      "learning_rate": 1.2579275243418074e-05,
+      "loss": 0.4503,
+      "step": 5780
+    },
+    {
+      "epoch": 0.6435149228891601,
+      "grad_norm": 1.21875,
+      "learning_rate": 1.245036409750725e-05,
+      "loss": 0.4796,
+      "step": 5800
+    },
+    {
+      "epoch": 0.64573393986464,
+      "grad_norm": 1.6015625,
+      "learning_rate": 1.2321646578020452e-05,
+      "loss": 0.4437,
+      "step": 5820
+    },
+    {
+      "epoch": 0.6479529568401198,
+      "grad_norm": 2.09375,
+      "learning_rate": 1.219313246012321e-05,
+      "loss": 0.4504,
+      "step": 5840
+    },
+    {
+      "epoch": 0.6501719738155997,
+      "grad_norm": 1.2265625,
+      "learning_rate": 1.2064831503534185e-05,
+      "loss": 0.5122,
+      "step": 5860
+    },
+    {
+      "epoch": 0.6523909907910795,
+      "grad_norm": 1.7890625,
+      "learning_rate": 1.1936753451783973e-05,
+      "loss": 0.4294,
+      "step": 5880
+    },
+    {
+      "epoch": 0.6546100077665594,
+      "grad_norm": 2.296875,
+      "learning_rate": 1.1808908031475151e-05,
+      "loss": 0.4895,
+      "step": 5900
+    },
+    {
+      "epoch": 0.6568290247420393,
+      "grad_norm": 1.140625,
+      "learning_rate": 1.1681304951543635e-05,
+      "loss": 0.4824,
+      "step": 5920
+    },
+    {
+      "epoch": 0.6590480417175192,
+      "grad_norm": 1.875,
+      "learning_rate": 1.1553953902521321e-05,
+      "loss": 0.525,
+      "step": 5940
+    },
+    {
+      "epoch": 0.661267058692999,
+      "grad_norm": 1.8359375,
+      "learning_rate": 1.1426864555800195e-05,
+      "loss": 0.4289,
+      "step": 5960
+    },
+    {
+      "epoch": 0.6634860756684788,
+      "grad_norm": 1.9375,
+      "learning_rate": 1.1300046562897837e-05,
+      "loss": 0.4695,
+      "step": 5980
+    },
+    {
+      "epoch": 0.6657050926439587,
+      "grad_norm": 1.7109375,
+      "learning_rate": 1.1173509554724461e-05,
+      "loss": 0.465,
+      "step": 6000
+    },
+    {
+      "epoch": 0.6679241096194386,
+      "grad_norm": 1.5625,
+      "learning_rate": 1.104726314085153e-05,
+      "loss": 0.4641,
+      "step": 6020
+    },
+    {
+      "epoch": 0.6701431265949185,
+      "grad_norm": 1.84375,
+      "learning_rate": 1.0921316908781965e-05,
+      "loss": 0.4559,
+      "step": 6040
+    },
+    {
+      "epoch": 0.6723621435703984,
+      "grad_norm": 1.4375,
+      "learning_rate": 1.079568042322205e-05,
+      "loss": 0.4938,
+      "step": 6060
+    },
+    {
+      "epoch": 0.6745811605458781,
+      "grad_norm": 1.2734375,
+      "learning_rate": 1.0670363225355054e-05,
+      "loss": 0.4247,
+      "step": 6080
+    },
+    {
+      "epoch": 0.676800177521358,
+      "grad_norm": 1.5390625,
+      "learning_rate": 1.0545374832116658e-05,
+      "loss": 0.4739,
+      "step": 6100
+    },
+    {
+      "epoch": 0.6790191944968379,
+      "grad_norm": 1.640625,
+      "learning_rate": 1.042072473547221e-05,
+      "loss": 0.4923,
+      "step": 6120
+    },
+    {
+      "epoch": 0.6812382114723178,
+      "grad_norm": 1.5,
+      "learning_rate": 1.0296422401695867e-05,
+      "loss": 0.5248,
+      "step": 6140
+    },
+    {
+      "epoch": 0.6834572284477977,
+      "grad_norm": 1.5390625,
+      "learning_rate": 1.017247727065172e-05,
+      "loss": 0.4336,
+      "step": 6160
+    },
+    {
+      "epoch": 0.6856762454232775,
+      "grad_norm": 1.7265625,
+      "learning_rate": 1.0048898755076885e-05,
+      "loss": 0.4915,
+      "step": 6180
+    },
+    {
+      "epoch": 0.6878952623987573,
+      "grad_norm": 1.515625,
+      "learning_rate": 9.925696239866679e-06,
+      "loss": 0.4908,
+      "step": 6200
+    },
+    {
+      "epoch": 0.6901142793742372,
+      "grad_norm": 1.4375,
+      "learning_rate": 9.802879081361927e-06,
+      "loss": 0.4512,
+      "step": 6220
+    },
+    {
+      "epoch": 0.6923332963497171,
+      "grad_norm": 2.21875,
+      "learning_rate": 9.680456606638376e-06,
+      "loss": 0.4356,
+      "step": 6240
+    },
+    {
+      "epoch": 0.694552313325197,
+      "grad_norm": 1.796875,
+      "learning_rate": 9.558438112798397e-06,
+      "loss": 0.4321,
+      "step": 6260
+    },
+    {
+      "epoch": 0.6967713303006768,
+      "grad_norm": 1.421875,
+      "learning_rate": 9.436832866264942e-06,
+      "loss": 0.4288,
+      "step": 6280
+    },
+    {
+      "epoch": 0.6989903472761566,
+      "grad_norm": 2.140625,
+      "learning_rate": 9.3156501020778e-06,
+      "loss": 0.4119,
+      "step": 6300
+    },
+    {
+      "epoch": 0.7012093642516365,
+      "grad_norm": 1.7421875,
+      "learning_rate": 9.194899023192295e-06,
+      "loss": 0.4729,
+      "step": 6320
+    },
+    {
+      "epoch": 0.7034283812271164,
+      "grad_norm": 2.03125,
+      "learning_rate": 9.074588799780359e-06,
+      "loss": 0.4438,
+      "step": 6340
+    },
+    {
+      "epoch": 0.7056473982025963,
+      "grad_norm": 1.703125,
+      "learning_rate": 8.95472856853414e-06,
+      "loss": 0.4509,
+      "step": 6360
+    },
+    {
+      "epoch": 0.7078664151780761,
+      "grad_norm": 1.453125,
+      "learning_rate": 8.835327431972136e-06,
+      "loss": 0.4812,
+      "step": 6380
+    },
+    {
+      "epoch": 0.710085432153556,
+      "grad_norm": 1.4375,
+      "learning_rate": 8.716394457747915e-06,
+      "loss": 0.4796,
+      "step": 6400
+    },
+    {
+      "epoch": 0.7123044491290358,
+      "grad_norm": 0.9375,
+      "learning_rate": 8.597938677961505e-06,
+      "loss": 0.4138,
+      "step": 6420
+    },
+    {
+      "epoch": 0.7145234661045157,
+      "grad_norm": 1.71875,
+      "learning_rate": 8.479969088473462e-06,
+      "loss": 0.4161,
+      "step": 6440
+    },
+    {
+      "epoch": 0.7167424830799956,
+      "grad_norm": 2.203125,
+      "learning_rate": 8.362494648221697e-06,
+      "loss": 0.4685,
+      "step": 6460
+    },
+    {
+      "epoch": 0.7189615000554754,
+      "grad_norm": 1.6875,
+      "learning_rate": 8.245524278541116e-06,
+      "loss": 0.4476,
+      "step": 6480
+    },
+    {
+      "epoch": 0.7211805170309553,
+      "grad_norm": 1.671875,
+      "learning_rate": 8.129066862486115e-06,
+      "loss": 0.5104,
+      "step": 6500
+    },
+    {
+      "epoch": 0.7233995340064352,
+      "grad_norm": 1.5625,
+      "learning_rate": 8.013131244155964e-06,
+      "loss": 0.4467,
+      "step": 6520
+    },
+    {
+      "epoch": 0.725618550981915,
+      "grad_norm": 2.234375,
+      "learning_rate": 7.89772622802316e-06,
+      "loss": 0.4317,
+      "step": 6540
+    },
+    {
+      "epoch": 0.7278375679573948,
+      "grad_norm": 2.109375,
+      "learning_rate": 7.782860578264806e-06,
+      "loss": 0.4398,
+      "step": 6560
+    },
+    {
+      "epoch": 0.7300565849328747,
+      "grad_norm": 1.6875,
+      "learning_rate": 7.668543018097014e-06,
+      "loss": 0.5054,
+      "step": 6580
+    },
+    {
+      "epoch": 0.7322756019083546,
+      "grad_norm": 1.4453125,
+      "learning_rate": 7.5547822291124715e-06,
+      "loss": 0.4968,
+      "step": 6600
+    },
+    {
+      "epoch": 0.7344946188838345,
+      "grad_norm": 1.5,
+      "learning_rate": 7.441586850621102e-06,
+      "loss": 0.4202,
+      "step": 6620
+    },
+    {
+      "epoch": 0.7367136358593144,
+      "grad_norm": 1.765625,
+      "learning_rate": 7.328965478993994e-06,
+      "loss": 0.4447,
+      "step": 6640
+    },
+    {
+      "epoch": 0.7389326528347941,
+      "grad_norm": 1.90625,
+      "learning_rate": 7.2169266670105555e-06,
+      "loss": 0.4869,
+      "step": 6660
+    },
+    {
+      "epoch": 0.741151669810274,
+      "grad_norm": 1.3203125,
+      "learning_rate": 7.105478923209001e-06,
+      "loss": 0.4744,
+      "step": 6680
+    },
+    {
+      "epoch": 0.7433706867857539,
+      "grad_norm": 0.76171875,
+      "learning_rate": 6.994630711240201e-06,
+      "loss": 0.4054,
+      "step": 6700
+    },
+    {
+      "epoch": 0.7455897037612338,
+      "grad_norm": 1.9765625,
+      "learning_rate": 6.884390449224898e-06,
+      "loss": 0.4307,
+      "step": 6720
+    },
+    {
+      "epoch": 0.7478087207367137,
+      "grad_norm": 1.921875,
+      "learning_rate": 6.774766509114435e-06,
+      "loss": 0.4728,
+      "step": 6740
+    },
+    {
+      "epoch": 0.7500277377121934,
+      "grad_norm": 1.53125,
+      "learning_rate": 6.66576721605496e-06,
+      "loss": 0.4254,
+      "step": 6760
+    },
+    {
+      "epoch": 0.7522467546876733,
+      "grad_norm": 1.78125,
+      "learning_rate": 6.557400847755183e-06,
+      "loss": 0.4508,
+      "step": 6780
+    },
+    {
+      "epoch": 0.7544657716631532,
+      "grad_norm": 1.828125,
+      "learning_rate": 6.449675633857772e-06,
+      "loss": 0.4814,
+      "step": 6800
+    },
+    {
+      "epoch": 0.7566847886386331,
+      "grad_norm": 1.578125,
+      "learning_rate": 6.3425997553143315e-06,
+      "loss": 0.459,
+      "step": 6820
+    },
+    {
+      "epoch": 0.758903805614113,
+      "grad_norm": 1.7421875,
+      "learning_rate": 6.236181343764144e-06,
+      "loss": 0.464,
+      "step": 6840
+    },
+    {
+      "epoch": 0.7611228225895929,
+      "grad_norm": 2.21875,
+      "learning_rate": 6.130428480916626e-06,
+      "loss": 0.4897,
+      "step": 6860
+    },
+    {
+      "epoch": 0.7633418395650726,
+      "grad_norm": 1.5625,
+      "learning_rate": 6.025349197937577e-06,
+      "loss": 0.4614,
+      "step": 6880
+    },
+    {
+      "epoch": 0.7655608565405525,
+      "grad_norm": 1.71875,
+      "learning_rate": 5.920951474839266e-06,
+      "loss": 0.444,
+      "step": 6900
+    },
+    {
+      "epoch": 0.7677798735160324,
+      "grad_norm": 1.1640625,
+      "learning_rate": 5.817243239874434e-06,
+      "loss": 0.4649,
+      "step": 6920
+    },
+    {
+      "epoch": 0.7699988904915123,
+      "grad_norm": 1.8046875,
+      "learning_rate": 5.714232368934163e-06,
+      "loss": 0.4758,
+      "step": 6940
+    },
+    {
+      "epoch": 0.7722179074669921,
+      "grad_norm": 2.15625,
+      "learning_rate": 5.611926684949779e-06,
+      "loss": 0.4519,
+      "step": 6960
+    },
+    {
+      "epoch": 0.774436924442472,
+      "grad_norm": 1.5234375,
+      "learning_rate": 5.510333957298756e-06,
+      "loss": 0.4729,
+      "step": 6980
+    },
+    {
+      "epoch": 0.7766559414179518,
+      "grad_norm": 1.28125,
+      "learning_rate": 5.409461901214679e-06,
+      "loss": 0.4557,
+      "step": 7000
+    },
+    {
+      "epoch": 0.7788749583934317,
+      "grad_norm": 1.2265625,
+      "learning_rate": 5.3093181772013545e-06,
+      "loss": 0.4433,
+      "step": 7020
+    },
+    {
+      "epoch": 0.7810939753689116,
+      "grad_norm": 1.4375,
+      "learning_rate": 5.209910390451007e-06,
+      "loss": 0.4767,
+      "step": 7040
+    },
+    {
+      "epoch": 0.7833129923443914,
+      "grad_norm": 2.140625,
+      "learning_rate": 5.111246090266763e-06,
+      "loss": 0.4422,
+      "step": 7060
+    },
+    {
+      "epoch": 0.7855320093198713,
+      "grad_norm": 1.7265625,
+      "learning_rate": 5.0133327694893035e-06,
+      "loss": 0.4276,
+      "step": 7080
+    },
+    {
+      "epoch": 0.7877510262953512,
+      "grad_norm": 1.5859375,
+      "learning_rate": 4.916177863927856e-06,
+      "loss": 0.482,
+      "step": 7100
+    },
+    {
+      "epoch": 0.789970043270831,
+      "grad_norm": 1.71875,
+      "learning_rate": 4.819788751795485e-06,
+      "loss": 0.4817,
+      "step": 7120
+    },
+    {
+      "epoch": 0.7921890602463109,
+      "grad_norm": 1.6484375,
+      "learning_rate": 4.7241727531487925e-06,
+      "loss": 0.4702,
+      "step": 7140
+    },
+    {
+      "epoch": 0.7944080772217907,
+      "grad_norm": 1.9296875,
+      "learning_rate": 4.629337129331983e-06,
+      "loss": 0.4494,
+      "step": 7160
+    },
+    {
+      "epoch": 0.7966270941972706,
+      "grad_norm": 1.9453125,
+      "learning_rate": 4.535289082425438e-06,
+      "loss": 0.4411,
+      "step": 7180
+    },
+    {
+      "epoch": 0.7988461111727505,
+      "grad_norm": 2.015625,
+      "learning_rate": 4.442035754698759e-06,
+      "loss": 0.4671,
+      "step": 7200
+    },
+    {
+      "epoch": 0.8010651281482304,
+      "grad_norm": 2.359375,
+      "learning_rate": 4.349584228068369e-06,
+      "loss": 0.4077,
+      "step": 7220
+    },
+    {
+      "epoch": 0.8032841451237102,
+      "grad_norm": 1.3984375,
+      "learning_rate": 4.257941523559703e-06,
+      "loss": 0.5378,
+      "step": 7240
+    },
+    {
+      "epoch": 0.80550316209919,
+      "grad_norm": 1.84375,
+      "learning_rate": 4.167114600773983e-06,
+      "loss": 0.4719,
+      "step": 7260
+    },
+    {
+      "epoch": 0.8077221790746699,
+      "grad_norm": 1.140625,
+      "learning_rate": 4.0771103573597125e-06,
+      "loss": 0.4596,
+      "step": 7280
+    },
+    {
+      "epoch": 0.8099411960501498,
+      "grad_norm": 1.5703125,
+      "learning_rate": 3.987935628488841e-06,
+      "loss": 0.4812,
+      "step": 7300
+    },
+    {
+      "epoch": 0.8121602130256297,
+      "grad_norm": 1.65625,
+      "learning_rate": 3.899597186337676e-06,
+      "loss": 0.4636,
+      "step": 7320
+    },
+    {
+      "epoch": 0.8143792300011095,
+      "grad_norm": 1.609375,
+      "learning_rate": 3.812101739572605e-06,
+      "loss": 0.4605,
+      "step": 7340
+    },
+    {
+      "epoch": 0.8165982469765893,
+      "grad_norm": 2.09375,
+      "learning_rate": 3.725455932840593e-06,
+      "loss": 0.4368,
+      "step": 7360
+    },
+    {
+      "epoch": 0.8188172639520692,
+      "grad_norm": 1.8125,
+      "learning_rate": 3.6396663462645917e-06,
+      "loss": 0.4686,
+      "step": 7380
+    },
+    {
+      "epoch": 0.8210362809275491,
+      "grad_norm": 1.875,
+      "learning_rate": 3.554739494943813e-06,
+      "loss": 0.4701,
+      "step": 7400
+    },
+    {
+      "epoch": 0.823255297903029,
+      "grad_norm": 1.8984375,
+      "learning_rate": 3.470681828458962e-06,
+      "loss": 0.4595,
+      "step": 7420
+    },
+    {
+      "epoch": 0.8254743148785089,
+      "grad_norm": 1.328125,
+      "learning_rate": 3.3874997303824416e-06,
+      "loss": 0.4265,
+      "step": 7440
+    },
+    {
+      "epoch": 0.8276933318539886,
+      "grad_norm": 1.5,
+      "learning_rate": 3.305199517793557e-06,
+      "loss": 0.4927,
+      "step": 7460
+    },
+    {
+      "epoch": 0.8299123488294685,
+      "grad_norm": 1.7109375,
+      "learning_rate": 3.2237874407987776e-06,
+      "loss": 0.4211,
+      "step": 7480
+    },
+    {
+      "epoch": 0.8321313658049484,
+      "grad_norm": 2.28125,
+      "learning_rate": 3.1432696820570993e-06,
+      "loss": 0.4771,
+      "step": 7500
+    },
+    {
+      "epoch": 0.8343503827804283,
+      "grad_norm": 1.7109375,
+      "learning_rate": 3.0636523563104985e-06,
+      "loss": 0.4934,
+      "step": 7520
+    },
+    {
+      "epoch": 0.8365693997559082,
+      "grad_norm": 1.6328125,
+      "learning_rate": 2.9849415099195886e-06,
+      "loss": 0.5184,
+      "step": 7540
+    },
+    {
+      "epoch": 0.838788416731388,
+      "grad_norm": 1.234375,
+      "learning_rate": 2.9071431204044123e-06,
+      "loss": 0.4622,
+      "step": 7560
+    },
+    {
+      "epoch": 0.8410074337068678,
+      "grad_norm": 1.6171875,
+      "learning_rate": 2.8302630959905084e-06,
+      "loss": 0.4491,
+      "step": 7580
+    },
+    {
+      "epoch": 0.8432264506823477,
+      "grad_norm": 2.46875,
+      "learning_rate": 2.7543072751602246e-06,
+      "loss": 0.4093,
+      "step": 7600
+    },
+    {
+      "epoch": 0.8454454676578276,
+      "grad_norm": 1.671875,
+      "learning_rate": 2.6792814262093214e-06,
+      "loss": 0.4312,
+      "step": 7620
+    },
+    {
+      "epoch": 0.8476644846333075,
+      "grad_norm": 1.71875,
+      "learning_rate": 2.605191246808912e-06,
+      "loss": 0.455,
+      "step": 7640
+    },
+    {
+      "epoch": 0.8498835016087873,
+      "grad_norm": 2.109375,
+      "learning_rate": 2.5320423635727824e-06,
+      "loss": 0.4839,
+      "step": 7660
+    },
+    {
+      "epoch": 0.8521025185842672,
+      "grad_norm": 2.15625,
+      "learning_rate": 2.45984033163006e-06,
+      "loss": 0.5647,
+      "step": 7680
+    },
+    {
+      "epoch": 0.854321535559747,
+      "grad_norm": 1.6484375,
+      "learning_rate": 2.388590634203366e-06,
+      "loss": 0.4108,
+      "step": 7700
+    },
+    {
+      "epoch": 0.8565405525352269,
+      "grad_norm": 1.90625,
+      "learning_rate": 2.3182986821923934e-06,
+      "loss": 0.3933,
+      "step": 7720
+    },
+    {
+      "epoch": 0.8587595695107068,
+      "grad_norm": 2.015625,
+      "learning_rate": 2.2489698137629904e-06,
+      "loss": 0.413,
+      "step": 7740
+    },
+    {
+      "epoch": 0.8609785864861866,
+      "grad_norm": 2.140625,
+      "learning_rate": 2.1806092939417732e-06,
+      "loss": 0.516,
+      "step": 7760
+    },
+    {
+      "epoch": 0.8631976034616665,
+      "grad_norm": 1.40625,
+      "learning_rate": 2.1132223142162714e-06,
+      "loss": 0.4403,
+      "step": 7780
+    },
+    {
+      "epoch": 0.8654166204371463,
+      "grad_norm": 1.859375,
+      "learning_rate": 2.046813992140679e-06,
+      "loss": 0.4355,
+      "step": 7800
+    },
+    {
+      "epoch": 0.8676356374126262,
+      "grad_norm": 1.4375,
+      "learning_rate": 1.981389370947218e-06,
+      "loss": 0.4744,
+      "step": 7820
+    },
+    {
+      "epoch": 0.869854654388106,
+      "grad_norm": 2.359375,
+      "learning_rate": 1.9169534191631243e-06,
+      "loss": 0.4106,
+      "step": 7840
+    },
+    {
+      "epoch": 0.8720736713635859,
+      "grad_norm": 2.046875,
+      "learning_rate": 1.853511030233354e-06,
+      "loss": 0.4519,
+      "step": 7860
+    },
+    {
+      "epoch": 0.8742926883390658,
+      "grad_norm": 1.7421875,
+      "learning_rate": 1.79106702214893e-06,
+      "loss": 0.4568,
+      "step": 7880
+    },
+    {
+      "epoch": 0.8765117053145457,
+      "grad_norm": 1.5234375,
+      "learning_rate": 1.7296261370810695e-06,
+      "loss": 0.4734,
+      "step": 7900
+    },
+    {
+      "epoch": 0.8787307222900255,
+      "grad_norm": 1.828125,
+      "learning_rate": 1.669193041021041e-06,
+      "loss": 0.547,
+      "step": 7920
+    },
+    {
+      "epoch": 0.8809497392655053,
+      "grad_norm": 1.4453125,
+      "learning_rate": 1.6097723234258188e-06,
+      "loss": 0.5001,
+      "step": 7940
+    },
+    {
+      "epoch": 0.8831687562409852,
+      "grad_norm": 1.6875,
+      "learning_rate": 1.5513684968695574e-06,
+      "loss": 0.447,
+      "step": 7960
+    },
+    {
+      "epoch": 0.8853877732164651,
+      "grad_norm": 1.578125,
+      "learning_rate": 1.4939859967008768e-06,
+      "loss": 0.4161,
+      "step": 7980
+    },
+    {
+      "epoch": 0.887606790191945,
+      "grad_norm": 1.78125,
+      "learning_rate": 1.437629180706037e-06,
+      "loss": 0.4606,
+      "step": 8000
+    },
+    {
+      "epoch": 0.8898258071674249,
+      "grad_norm": 1.6015625,
+      "learning_rate": 1.382302328778e-06,
+      "loss": 0.4126,
+      "step": 8020
+    },
+    {
+      "epoch": 0.8920448241429046,
+      "grad_norm": 1.6015625,
+      "learning_rate": 1.328009642591394e-06,
+      "loss": 0.4313,
+      "step": 8040
+    },
+    {
+      "epoch": 0.8942638411183845,
+      "grad_norm": 1.4296875,
+      "learning_rate": 1.2747552452834388e-06,
+      "loss": 0.4436,
+      "step": 8060
+    },
+    {
+      "epoch": 0.8964828580938644,
+      "grad_norm": 1.6015625,
+      "learning_rate": 1.2225431811408133e-06,
+      "loss": 0.4591,
+      "step": 8080
+    },
+    {
+      "epoch": 0.8987018750693443,
+      "grad_norm": 1.484375,
+      "learning_rate": 1.1713774152925195e-06,
+      "loss": 0.4103,
+      "step": 8100
+    },
+    {
+      "epoch": 0.9009208920448242,
+      "grad_norm": 1.890625,
+      "learning_rate": 1.1212618334087693e-06,
+      "loss": 0.4196,
+      "step": 8120
+    },
+    {
+      "epoch": 0.903139909020304,
+      "grad_norm": 2.125,
+      "learning_rate": 1.0722002414058868e-06,
+      "loss": 0.4701,
+      "step": 8140
+    },
+    {
+      "epoch": 0.9053589259957838,
+      "grad_norm": 1.8671875,
+      "learning_rate": 1.0241963651572867e-06,
+      "loss": 0.4452,
+      "step": 8160
+    },
+    {
+      "epoch": 0.9075779429712637,
+      "grad_norm": 2.21875,
+      "learning_rate": 9.772538502105093e-07,
+      "loss": 0.4122,
+      "step": 8180
+    },
+    {
+      "epoch": 0.9097969599467436,
+      "grad_norm": 1.625,
+      "learning_rate": 9.313762615103761e-07,
+      "loss": 0.4265,
+      "step": 8200
+    },
+    {
+      "epoch": 0.9120159769222235,
+      "grad_norm": 1.9296875,
+      "learning_rate": 8.865670831282513e-07,
+      "loss": 0.4678,
+      "step": 8220
+    },
+    {
+      "epoch": 0.9142349938977034,
+      "grad_norm": 1.671875,
+      "learning_rate": 8.42829717997457e-07,
+      "loss": 0.4137,
+      "step": 8240
+    },
+    {
+      "epoch": 0.9164540108731832,
+      "grad_norm": 1.6484375,
+      "learning_rate": 8.001674876548471e-07,
+      "loss": 0.4939,
+      "step": 8260
+    },
+    {
+      "epoch": 0.918673027848663,
+      "grad_norm": 1.640625,
+      "learning_rate": 7.585836319885525e-07,
+      "loss": 0.4888,
+      "step": 8280
+    },
+    {
+      "epoch": 0.9208920448241429,
+      "grad_norm": 1.578125,
+      "learning_rate": 7.180813089919403e-07,
+      "loss": 0.4352,
+      "step": 8300
+    },
+    {
+      "epoch": 0.9231110617996228,
+      "grad_norm": 1.6640625,
+      "learning_rate": 6.78663594523788e-07,
+      "loss": 0.5086,
+      "step": 8320
+    },
+    {
+      "epoch": 0.9253300787751026,
+      "grad_norm": 1.96875,
+      "learning_rate": 6.403334820746876e-07,
+      "loss": 0.4695,
+      "step": 8340
+    },
+    {
+      "epoch": 0.9275490957505825,
+      "grad_norm": 1.5234375,
+      "learning_rate": 6.030938825397225e-07,
+      "loss": 0.3946,
+      "step": 8360
+    },
+    {
+      "epoch": 0.9297681127260623,
+      "grad_norm": 1.609375,
+      "learning_rate": 5.669476239973975e-07,
+      "loss": 0.4797,
+      "step": 8380
+    },
+    {
+      "epoch": 0.9319871297015422,
+      "grad_norm": 1.796875,
+      "learning_rate": 5.318974514948672e-07,
+      "loss": 0.4274,
+      "step": 8400
+    },
+    {
+      "epoch": 0.9342061466770221,
+      "grad_norm": 1.453125,
+      "learning_rate": 4.979460268394726e-07,
+      "loss": 0.4369,
+      "step": 8420
+    },
+    {
+      "epoch": 0.936425163652502,
+      "grad_norm": 1.6640625,
+      "learning_rate": 4.6509592839659666e-07,
+      "loss": 0.4452,
+      "step": 8440
+    },
+    {
+      "epoch": 0.9386441806279818,
+      "grad_norm": 1.7734375,
+      "learning_rate": 4.333496508938506e-07,
+      "loss": 0.4949,
+      "step": 8460
+    },
+    {
+      "epoch": 0.9408631976034617,
+      "grad_norm": 2.078125,
+      "learning_rate": 4.02709605231627e-07,
+      "loss": 0.4815,
+      "step": 8480
+    },
+    {
+      "epoch": 0.9430822145789415,
+      "grad_norm": 2.109375,
+      "learning_rate": 3.731781182999983e-07,
+      "loss": 0.5533,
+      "step": 8500
+    },
+    {
+      "epoch": 0.9453012315544214,
+      "grad_norm": 1.5234375,
+      "learning_rate": 3.447574328020109e-07,
+      "loss": 0.4741,
+      "step": 8520
+    },
+    {
+      "epoch": 0.9475202485299012,
+      "grad_norm": 1.5546875,
+      "learning_rate": 3.1744970708337205e-07,
+      "loss": 0.5182,
+      "step": 8540
+    },
+    {
+      "epoch": 0.9497392655053811,
+      "grad_norm": 1.4296875,
+      "learning_rate": 2.912570149685323e-07,
+      "loss": 0.4612,
+      "step": 8560
+    },
+    {
+      "epoch": 0.951958282480861,
+      "grad_norm": 1.3984375,
+      "learning_rate": 2.661813456032014e-07,
+      "loss": 0.4253,
+      "step": 8580
+    },
+    {
+      "epoch": 0.9541772994563409,
+      "grad_norm": 2.125,
+      "learning_rate": 2.4222460330327933e-07,
+      "loss": 0.5029,
+      "step": 8600
+    },
+    {
+      "epoch": 0.9563963164318207,
+      "grad_norm": 1.9921875,
+      "learning_rate": 2.1938860741023858e-07,
+      "loss": 0.4444,
+      "step": 8620
+    },
+    {
+      "epoch": 0.9586153334073005,
+      "grad_norm": 1.390625,
+      "learning_rate": 1.9767509215296297e-07,
+      "loss": 0.4607,
+      "step": 8640
+    },
+    {
+      "epoch": 0.9608343503827804,
+      "grad_norm": 1.7890625,
+      "learning_rate": 1.7708570651604306e-07,
+      "loss": 0.5045,
+      "step": 8660
+    },
+    {
+      "epoch": 0.9630533673582603,
+      "grad_norm": 1.484375,
+      "learning_rate": 1.5762201411454626e-07,
+      "loss": 0.4525,
+      "step": 8680
+    },
+    {
+      "epoch": 0.9652723843337402,
+      "grad_norm": 1.28125,
+      "learning_rate": 1.3928549307527183e-07,
+      "loss": 0.4484,
+      "step": 8700
+    },
+    {
+      "epoch": 0.9674914013092201,
+      "grad_norm": 1.7265625,
+      "learning_rate": 1.2207753592450078e-07,
+      "loss": 0.4194,
+      "step": 8720
+    },
+    {
+      "epoch": 0.9697104182846998,
+      "grad_norm": 1.7734375,
+      "learning_rate": 1.059994494822386e-07,
+      "loss": 0.4822,
+      "step": 8740
+    },
+    {
+      "epoch": 0.9719294352601797,
+      "grad_norm": 1.7421875,
+      "learning_rate": 9.105245476297653e-08,
+      "loss": 0.4584,
+      "step": 8760
+    },
+    {
+      "epoch": 0.9741484522356596,
+      "grad_norm": 1.5546875,
+      "learning_rate": 7.723768688296217e-08,
+      "loss": 0.4508,
+      "step": 8780
+    },
+    {
+      "epoch": 0.9763674692111395,
+      "grad_norm": 1.4921875,
+      "learning_rate": 6.455619497399534e-08,
+      "loss": 0.4953,
+      "step": 8800
+    },
+    {
+      "epoch": 0.9785864861866194,
+      "grad_norm": 1.9375,
+      "learning_rate": 5.300894210375329e-08,
+      "loss": 0.4791,
+      "step": 8820
+    },
+    {
+      "epoch": 0.9808055031620991,
+      "grad_norm": 1.4765625,
+      "learning_rate": 4.259680520265596e-08,
+      "loss": 0.4083,
+      "step": 8840
+    },
+    {
+      "epoch": 0.983024520137579,
+      "grad_norm": 1.6328125,
+      "learning_rate": 3.3320574997267595e-08,
+      "loss": 0.449,
+      "step": 8860
+    },
+    {
+      "epoch": 0.9852435371130589,
+      "grad_norm": 1.84375,
+      "learning_rate": 2.5180955950243056e-08,
+      "loss": 0.5022,
+      "step": 8880
+    },
+    {
+      "epoch": 0.9874625540885388,
+      "grad_norm": 1.203125,
+      "learning_rate": 1.8178566206837334e-08,
+      "loss": 0.54,
+      "step": 8900
+    },
+    {
+      "epoch": 0.9896815710640187,
+      "grad_norm": 1.8203125,
+      "learning_rate": 1.231393754795307e-08,
+      "loss": 0.4788,
+      "step": 8920
+    },
+    {
+      "epoch": 0.9919005880394985,
+      "grad_norm": 1.5546875,
+      "learning_rate": 7.587515349762874e-09,
+      "loss": 0.5039,
+      "step": 8940
+    },
+    {
+      "epoch": 0.9941196050149783,
+      "grad_norm": 1.8984375,
+      "learning_rate": 3.9996585498797145e-09,
+      "loss": 0.5317,
+      "step": 8960
+    },
+    {
+      "epoch": 0.9963386219904582,
+      "grad_norm": 2.0,
+      "learning_rate": 1.550639620103711e-09,
+      "loss": 0.4578,
+      "step": 8980
+    },
+    {
+      "epoch": 0.9985576389659381,
+      "grad_norm": 1.5234375,
+      "learning_rate": 2.406445457253659e-10,
+      "loss": 0.474,
+      "step": 9000
+    },
+    {
+      "epoch": 1.0,
+      "step": 9013,
+      "total_flos": 5.485972481640161e+17,
+      "train_loss": 0.4824531834622566,
+      "train_runtime": 26326.043,
+      "train_samples_per_second": 0.685,
+      "train_steps_per_second": 0.342
+    }
+  ],
+  "logging_steps": 20,
+  "max_steps": 9013,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 100,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.485972481640161e+17,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}