llama3-8b-Lawyer

LLaMA3-8B-Lawyer

This project involves fine-tuning the LLaMA3-8B model using the dzunggg/legal-qa-v1 dataset. The fine-tuning was conducted with the LLaMA-Factory toolkit on a single NVIDIA L20-48G GPU. The fine-tuned model has been uploaded to Hugging Face and is available at StevenChen16/llama3-8b-Lawyer.

Project Overview

The primary goal of this project was to create a high-performing legal question-answering model based on LLaMA3-8B. By leveraging the dzunggg/legal-qa-v1 dataset and the capabilities of LLaMA-Factory, we were able to fine-tune the model effectively. The AI model can function like a lawyer, asking detailed questions about the case background and making judgments based on the provided information.

Fine-Tuning Details

Model

• Base Model: nvidia/Llama3-ChatQA-1.5-8B
• Fine-Tuned Model: StevenChen16/llama3-8b-Lawyer

Dataset

• Dataset Used: dzunggg/legal-qa-v1

Training Configuration

args = dict(
  stage="sft",                        # do supervised fine-tuning
  do_train=True,
  model_name_or_path="nvidia/Llama3-ChatQA-1.5-8B", # use bnb-4bit-quantized Llama-3-8B-Instruct model
  dataset="legal_qa_v1_train",             # use legal_qa_v1_train dataset
  template="llama3",                     # use llama3 prompt template
  finetuning_type="lora",                   # use LoRA adapters to save memory
  lora_target="all",                     # attach LoRA adapters to all linear layers
  output_dir="llama3_lora",                  # the path to save LoRA adapters
  per_device_train_batch_size=8,               # the batch size
  gradient_accumulation_steps=6,               # the gradient accumulation steps
  lr_scheduler_type="cosine",                 # use cosine learning rate scheduler
  logging_steps=10,                      # log every 10 steps
  warmup_ratio=0.1,                      # use warmup scheduler
  save_steps=1000,                      # save checkpoint every 1000 steps
  learning_rate=1e-4,                     # the learning rate
  num_train_epochs=10.0,                    # the epochs of training
  max_samples=500,                      # use 500 examples in each dataset
  max_grad_norm=1.0,                     # clip gradient norm to 1.0
  quantization_bit=8,                     # use 8-bit quantization
  loraplus_lr_ratio=16.0,                   # use LoRA+ algorithm with lambda=16.0
  use_unsloth=True,                      # use UnslothAI's LoRA optimization for 2x faster training
  fp16=True,                         # use float16 mixed precision training
  overwrite_output_dir=True,
)

Hardware

• GPU: NVIDIA L20-48G

Usage

You can load and use the fine-tuned model from Hugging Face as follows:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "StevenChen16/llama3-8b-Lawyer"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Example usage
input_text = "Your legal question here."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Example Interaction

The model can engage in a detailed interaction, simulating the behavior of a lawyer. Provide the case background, and the model will ask for more details to make informed judgments.

Example

input_text = "I have a contract dispute where the other party did not deliver the promised goods."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Output:

Can you provide more details about the contract terms and the goods that were supposed to be delivered? Were there any specific deadlines mentioned in the contract?

Training Notebook and Repository

• Training Notebook: Google Colab Notebook
• GitHub Repository: lawyer-llama3-8b

Results

The fine-tuned model has shown promising results in understanding and answering legal questions. By leveraging advanced techniques such as LoRA and UnslothAI optimizations, the training process was efficient and effective, ensuring a high-quality model output.

Acknowledgements

• LLaMA-Factory
• Dataset: dzunggg/legal-qa-v1
• Base Model: nvidia/Llama3-ChatQA-1.5-8B
• Hosted on Hugging Face

License

This project is licensed under the MIT License. See the LICENSE file for details.

---