Call of Cthulhu (CoC) Dialogue Generation

This project utilizes fine-tuned language models to generate dialogues based on the "Call of Cthulhu" role-playing game.

Installation Steps

1. Create and activate a new Conda environment:

   conda create -n coc python=3.8
   conda activate coc

2. Install required dependencies:

   pip install -r requirements.txt

3. Set up Accelerate with DeepSpeed:

   accelerate config

Configuration

DeepSpeed Settings

Adjust the DeepSpeed settings to meet your requirements by modifying the configuration file located at config/deepspeed.json.

Dataset

CoC Dialogue Dataset

The dataset is located at data/coc_dialogue.json and is specifically crafted for fine-tuning the language model on CoC dialogue generation.

Dataset Generation

The folder coc_data_gen/ details the process of dataset generation using the GPT-3.5-turbo API.

Model Weights

Download the supervised fine-tuned weights for Baichuan-7B and Llama2-7B from the following link:

• Baichuan-7B and Llama2-7B SFT Weights

Training

To train the language model, use the following command which leverages Accelerate with DeepSpeed:

ACCELERATE_USE_DEEPSPEED=true CUDA_VISIBLE_DEVICES="0,1" accelerate launch finetuning.py

Adjust the CUDA_VISIBLE_DEVICES parameter as necessary to specify the GPUs you intend to use.

Inference

Generate dialogues with the trained model by running:

python inference.py

This script takes prompts and generates dialogues using the fine-tuned model.

Directory Structure

• coc_data_gen/: Scripts for CoC dataset generation using the GPT-3.5-turbo API.
• config/: DeepSpeed configuration file.
• data/: CoC dialogue dataset and additional evaluation data.
• samples/: Sample prompts for inference.
• finetuning.py: Training script for the language model.
• inference.py: Inference script for dialogue generation.
• requirements.txt: Required Python dependencies.

License

This project is released under the MIT License.