Intel/neural-chat-7b-v3-1

Model Details: Neural-Chat-v3-1

This model is a fine-tuned 7B parameter LLM on the Intel Gaudi 2 processor from the mistralai/Mistral-7B-v0.1 on the open source dataset Open-Orca/SlimOrca. The model was aligned using the Direct Performance Optimization (DPO) method with Intel/orca_dpo_pairs. For more information, refer to the Medium article The Practice of Supervised Fine-tuning and Direct Preference Optimization on Intel Gaudi2.

Photo by Google DeepMind on Unsplash

Model Detail	Description
Model Authors - Company	Intel. The NeuralChat team with members from DCAI/AISE/AIPT. Core team members: Kaokao Lv, Liang Lv, Chang Wang, Wenxin Zhang, Xuhui Ren, and Haihao Shen.
Date	October, 2023
Version	v3-1
Type	7B Large Language Model
Paper or Other Resources	Medium Blog
License	Apache 2.0
Questions or Comments	Community Tab and Intel DevHub Discord

Intended Use	Description
Primary intended uses	You can use the fine-tuned model for several language-related tasks. Checkout the LLM Leaderboard to see how this model is doing.
Primary intended users	Anyone doing inference on language-related tasks.
Out-of-scope uses	This model in most cases will need to be fine-tuned for your particular task. The model should not be used to intentionally create hostile or alienating environments for people.

How To Use

Context length for this model: 8192 tokens (same as https://huggingface.co/mistralai/Mistral-7B-v0.1)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-04
• train_batch_size: 1
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-HPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 2.0

Reproduce the model

Here is the sample code to reproduce the model: GitHub sample code. Here is the documentation to reproduce building the model:

git clone https://github.com/intel/intel-extension-for-transformers.git
cd intel-extension-for-transformers

docker build --no-cache ./ --target hpu --build-arg REPO=https://github.com/intel/intel-extension-for-transformers.git --build-arg ITREX_VER=main -f ./intel_extension_for_transformers/neural_chat/docker/Dockerfile -t chatbot_finetuning:latest

docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host chatbot_finetuning:latest

# after entering docker container
cd examples/finetuning/finetune_neuralchat_v3

We select the latest pretrained mistralai/Mistral-7B-v0.1 and the open source dataset Open-Orca/SlimOrca to conduct the experiment.

The below script use deepspeed zero2 to lanuch the training with 8 cards Gaudi2. In the finetune_neuralchat_v3.py, the default use_habana=True, use_lazy_mode=True, device="hpu" for Gaudi2. And if you want to run it on NVIDIA GPU, you can set them use_habana=False, use_lazy_mode=False, device="auto".

deepspeed --include localhost:0,1,2,3,4,5,6,7 \
    --master_port 29501 \
    finetune_neuralchat_v3.py

Merge the LoRA weights:

python apply_lora.py \
    --base-model-path mistralai/Mistral-7B-v0.1 \
    --lora-model-path finetuned_model/ \
    --output-path finetuned_model_lora

FP32 Inference with Transformers

import transformers

model_name = 'Intel/neural-chat-7b-v3-1'
model = transformers.AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)

def generate_response(system_input, user_input):

    # Format the input using the provided template
    prompt = f"### System:\n{system_input}\n### User:\n{user_input}\n### Assistant:\n"

    # Tokenize and encode the prompt
    inputs = tokenizer.encode(prompt, return_tensors="pt", add_special_tokens=False)

    # Generate a response
    outputs = model.generate(inputs, max_length=1000, num_return_sequences=1)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)

    # Extract only the assistant's response
    return response.split("### Assistant:\n")[-1]


# Example usage
system_input = "You are a math expert assistant. Your mission is to help users understand and solve various math problems. You should provide step-by-step solutions, explain reasonings and give the correct answer."
user_input = "calculate 100 + 520 + 60"
response = generate_response(system_input, user_input)
print(response)

# expected response
"""
To calculate the sum of 100, 520, and 60, we will follow these steps:

1. Add the first two numbers: 100 + 520
2. Add the result from step 1 to the third number: (100 + 520) + 60

Step 1: Add 100 and 520
100 + 520 = 620

Step 2: Add the result from step 1 to the third number (60)
(620) + 60 = 680

So, the sum of 100, 520, and 60 is 680.
"""

BF16 Inference with Intel Extension for Transformers and Intel Extension for Pytorch

from transformers import AutoTokenizer, TextStreamer
import torch
from intel_extension_for_transformers.transformers import AutoModelForCausalLM
import intel_extension_for_pytorch as ipex

model_name = "Intel/neural-chat-7b-v3-1"
prompt = "Once upon a time, there existed a little girl,"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
inputs = tokenizer(prompt, return_tensors="pt").input_ids
streamer = TextStreamer(tokenizer)

model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
model = ipex.optimize(model.eval(), dtype=torch.bfloat16, inplace=True, level="O1", auto_kernel_selection=True)

outputs = model.generate(inputs, streamer=streamer, max_new_tokens=300)

INT4 Inference with Transformers and Intel Extension for Transformers

from transformers import AutoTokenizer, TextStreamer
from intel_extension_for_transformers.transformers import AutoModelForCausalLM, WeightOnlyQuantConfig
model_name = "Intel/neural-chat-7b-v3-1"

# for int8, should set weight_dtype="int8"       
config = WeightOnlyQuantConfig(compute_dtype="bf16", weight_dtype="int4")
prompt = "Once upon a time, there existed a little girl,"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
inputs = tokenizer(prompt, return_tensors="pt").input_ids
streamer = TextStreamer(tokenizer)

model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=config)
outputs = model.generate(inputs, streamer=streamer, max_new_tokens=300)

Factors	Description
Groups	More details about the dataset and annotations can be found at Open-Orca/SlimOrca and the associated paper at https://arxiv.org/abs/2306.02707.
Instrumentation	The performance of the model can vary depending on the inputs to the model. In this case, the prompts provided can drastically change the prediction of the language model.
Environment	The model was trained on the Intel Gaudi 2 processor (8 cards).
Card Prompts	Model deployment on alternate hardware and software will change model performance. The model evaluation factors are from the Hugging Face LLM leaderboard: ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K, and DROP (see Quantitative Analyses below).

Metrics	Description
Model performance measures	The model performance was evaluated against other LLMs according to the measures on the LLM leaderboard. These were selected as this has become the standard for LLM performance.
Decision thresholds	No decision thresholds were used.
Approaches to uncertainty and variability	-

Training and Evaluation Data	Description
Datasets	The training data are from Open-Orca/SlimOrca. There is no contamination from the GSM8k test set, as this is not a part of the Open-Orca/SlimOrca dataset.
Motivation	-
Preprocessing	-

Quantitative Analyses

The model was submitted to the LLM Leaderboard. The detailed submission can be found here: https://huggingface.co/datasets/open-llm-leaderboard/details_Intel__neural-chat-7b-v3-1. The metrics can be found below and show that the model has significantly improved performance from Mistral-7B-v0.1 and neural-chat-7b-v3.

Model	Average ⬆️	ARC (25-s) ⬆️	HellaSwag (10-s) ⬆️	MMLU (5-s) ⬆️	TruthfulQA (MC) (0-s) ⬆️	Winogrande (5-s)	GSM8K (5-s)	DROP (3-s)
mistralai/Mistral-7B-v0.1	50.32	59.58	83.31	64.16	42.15	78.37	18.12	6.14
Intel/neural-chat-7b-v3	57.31	67.15	83.29	62.26	58.77	78.06	1.21	50.43
Intel/neural-chat-7b-v3-1	59.06	66.21	83.64	62.37	59.65	78.14	19.56	43.84

Testing Model Quantizability

The following code block can be run to determine, for PyTorch models, if that model is amenable to quantization.
One caveat - the Intel Extension for PyTorch uses optimum ipex, which is pre-release and needs further testing.

To install the dependencies, you should first install Intel Extensions for PyTorch and tehn pip install each of the following dependencies:

• torch
• optimum.intel
• optimum[ipex]
• transformers

Intel Extension for PyTorch method:

In this case, we are testing if neural-chat-7b-v3-1 can be quantized and this testing method demonstrates the model size change, for example: when the base type is specified to be torch.bfloat16 but also specifying that load_in_4bit=True which causes the weights only to be quantized we see an output from the model testing as follows:

• model_quantize_internal: model size = 27625.02 MB
• model_quantize_internal: quant size = 4330.80 MB

This code should run from within a python script - such as ipex_test.py as follows:

import torch
import os
from transformers import AutoTokenizer
from intel_extension_for_transformers.transformers import AutoModelForCausalLM, pipeline
model_name = "Intel/neural-chat-7b-v3-1"     
prompt = "Once upon a time, there existed a little girl,"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
inputs = tokenizer(prompt, return_tensors="pt").input_ids

result = {torch.bfloat16:"failed"}
typ = torch.bfloat16
try:
    model = AutoModelForCausalLM.from_pretrained(model_name, load_in_4bit=True,  torch_dtype = typ)
    outputs = model.generate(inputs, max_new_tokens=20)
    result[typ] = f"passed, {os.stat(model.bin_file).st_size}"
except:
    result[typ] = "failed"

    
print("\n\nResults of quantizing: ")  
# determine if Quantized
with open(r"output.log", 'r') as fp:
    for l_no, line in enumerate(fp):
        # search string
        if 'model_quantize_internal' in line:
            print(line)
            
print("\n\nExecution results ")
for k,v in result.items():
    print(k,v)
    
print("\n\nModel Output: ")
tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

Run the code as folows from a bash terminal:

python ipex_test.py 2>&1 | tee output.log

The entire output is captured in the output.log but it will be summarized, along with output from the model indicating either pass or fail of the quantization as well as model output for a given prompt.

Ethical Considerations and Limitations

Neural-chat-7b-v3-1 can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Therefore, before deploying any applications of neural-chat-7b-v3-1, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

Here are a couple of useful links to learn more about Intel's AI software:

• Intel Neural Compressor link
• Intel Extension for Transformers link

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.