Classical-Vernacular Chinese Translation

This project aims to translate between classical and vernacular Chinese. We applied the Quantized Low Rank Adaptation (QLoRA) on the Taiwan-LLM v2.0 7B model to perform parameter-efficient fine-tuning (PEFT), which largely decrease computational and storage costs.

Prerequisites

• Python 3.10

Repository Structure

• requirements.txt: List of Python packages required for this project.
• download.sh: Bash script to download training data and a pre-trained PEFT adapter.
• run.sh: Bash script to run the inference code.
• src: Folder containing resources used by shell scripts and for training models on your own.
• report.pdf: Explanations of the hyperparameter sweep and generation strategies.
• ppl.py: Script to calculate the model perplexity.

Setup

Step 1: Clone the Repository

git clone https://github.com/Hannibal0420/Classical-Vernacular-Chinese-Translation.git
cd Classical-Vernacular-Chinese-Translation

Step 2: Install Dependencies

Install the required Python packages and inference model using the following command:

pip install -r requirements.txt
bash ./download.sh

Run Inference

To run the inference code, execute the run.sh script with the following arguments:

• ${1}: Path to the Taiwan-LLaMa checkpoint folder
• ${2}: Path to the folder containing the peft model downloaded by download.sh
• ${3}: Path to the input file (.json)
• ${4}: Path to the output file (.json)

Note: Make sure to replace /path/to/input.json and /path/to/output.json with the actual paths to your files. To use the example code in this repo, you can run as below.

bash ./run.sh ./Taiwan-LLM-7B-v2.0-chat ./adapter_checkpoint ./data/private_test.json ./output.json

Train Your Own Model

In the train_src folder, you can fine-tune existing models or train from scratch. These codes are modified from thess sources:

• Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA (article)
• QLoRA: Efficient Finetuning of Quantized LLMs (repo)
  • Finetuning notebook
  • Inference notebook

License

This project is licensed under the MIT License - see the LICENSE.md file for details.