llm-jp-3-13b-finetune

LLM Fine-Tuning and Inference Pipeline

Overview

This repository contains a pipeline for fine-tuning and inference of a large language model (LLM) using Hugging Face's transformers library and PEFT (Parameter-Efficient Fine-Tuning) techniques such as LoRA. The pipeline is designed to handle Japanese instruction datasets and efficiently adapt large models to specific tasks.

Key Features

• Model Preparation: Load and configure a pretrained causal language model for fine-tuning.
• LoRA Integration: Apply Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning.
• Custom Dataset Handling: Supports JSON and JSONL formats for instruction datasets.
• Prompt Formatting: Customizable prompt templates for instruction-following tasks.
• Inference Pipeline: Easily run tasks and generate outputs with the fine-tuned model.

---

How to Use

1. Model Preparation

Ensure you have access to a pretrained LLM. The model ID or path can be specified in the base_model_id variable. If the model is stored locally, update the path accordingly.

2. Dataset Preparation

Prepare your dataset in JSON or JSONL format. For example:

{
    "task_id": "1",
    "input": "Translate this text to Japanese.",
    "output": "このテキストを日本語に翻訳してください。"
}

Save the dataset in the appropriate directory, and provide the path in the load_dataset function.

3. Fine-Tuning

Run the fine-tuning script with your dataset. Key configurations include:

• LoRA Parameters: Adjustable rank (r), scaling factor (lora_alpha), and dropout rate (lora_dropout).
• Training Arguments: Includes batch size, learning rate, gradient accumulation steps, and evaluation strategy.

4. Inference

Use the inference script to process new tasks. Customize the prompt template to fit your specific use case. Results are saved in JSONL format for easy sharing and analysis.

---

Fine-Tuning Configuration

The fine-tuning process leverages the following parameters:

• LoRA: Rank (r), scaling factor (lora_alpha), and dropout rate.
• Training:
  • Batch size per device.
  • Gradient accumulation steps.
  • Learning rate scheduler and optimizer.

---

Example Workflow

1. Clone the repository and navigate to the directory.
2. Place your dataset in the specified path.
3. Configure base_model_id and new_model_id.
4. Run the fine-tuning script:

   python fine_tune.py
   

5. Perform inference with:

   python inference.py
   

---

License

• The datasets used in this repository are subject to the CC-BY-NC-SA license. Ensure compliance with the licensing terms before usage.
• Models fine-tuned using this repository inherit the dataset's license.

---

Citation

If you use this repository or the associated datasets, please cite the following paper:

関根聡, 安藤まや, 後藤美知子, 鈴木久美, 河原大輔, 井之上直也, 乾健太郎. Ichikara Instruction: LLMのための日本語インストラクションデータの構築. 言語処理学会第30回年次大会(2024).

---

Acknowledgments

This repository was developed with the support of Hugging Face, PEFT, and LoRA technologies.