See axolotl config

axolotl version: 0.4.1

base_model: microsoft/Phi-3.5-mini-instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
chat_template: phi_3

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: flydust/CodeGen_50000_Test
    type: chat_template
    field_messages: conversations
    # The key in the message turn that contains the role. Default is "role".
    message_field_role: from
    # The key in the message turn that contains the content. Default is "content".
    message_field_content: value
    # Optional[Dict[str, List]]. Roles mapping for the messages.
    roles:
      user: ["human", "user"]
      assistant: ["gpt", "assistant", "ai"]
      system: ["system"]


dataset_prepared_path: last_run_prepared
val_set_size: 0.001
output_dir: axolotl_out/Phi-3.5-mini-instruct-Code50000-Test

sequence_len: 4096
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project: SynDa
wandb_entity:
wandb_watch:
wandb_name: Phi-3.5-mini-instruct-Code50000-Test
wandb_log_model:
hub_model_id: flydust/Phi-3.5-mini-instruct-Code50000-Test

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

train_on_inputs: false
group_by_length: false
bf16: true
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
# Disable flash attention
flash_attention: true
# sdp_attention: falses
# eager_attention: true

warmup_ratio: 0.1
evals_per_epoch: 10
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Phi-3.5-mini-instruct-Code50000-Test

This model is a fine-tuned version of microsoft/Phi-3.5-mini-instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1712

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: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 21
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
0.4415	0.0093	1	0.4636
0.2376	0.1019	11	0.2367
0.2014	0.2037	22	0.2002
0.1824	0.3056	33	0.1895
0.1728	0.4074	44	0.1817
0.1764	0.5093	55	0.1786
0.1822	0.6111	66	0.1766
0.1661	0.7130	77	0.1750
0.171	0.8148	88	0.1740
0.1577	0.9167	99	0.1741
0.1615	1.0162	110	0.1722
0.1551	1.1181	121	0.1720
0.1676	1.2199	132	0.1724
0.1583	1.3218	143	0.1714
0.164	1.4236	154	0.1713
0.1581	1.5255	165	0.1717
0.1496	1.6273	176	0.1707
0.1563	1.7292	187	0.1710
0.1518	1.8310	198	0.1707
0.1687	1.9329	209	0.1712

Framework versions

• Transformers 4.44.2
• Pytorch 2.5.1+cu124
• Datasets 2.20.0
• Tokenizers 0.19.1