INC4AI/neural-chat-7b-v3

Model Details: Neural-Chat-v3

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

How To Use

Context length for this model: 8192 tokens (same as 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:
• 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

You can then align the model following the steps in the GitHub sample code.

Use the model

FP32 Inference with Transformers

import transformers


model_name = 'Intel/neural-chat-7b-v3'
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"
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"

# for int8, should set weight_dtype="int8"      
config = WeightOnlyQuantConfig(compute_dtype="bf16", weight_dtype="int4")
prompt = "Once upon a time, there was a horse in the forest,"

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)

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. The metrics can be found below and show that the model has significantly improved performance from Mistral-7B-v0.1.

Ethical Considerations and Limitations

Neural-chat-7b-v3 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 the model, 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.