This repository provides resources to fine-tune the Llama-7B-Uncensored model using the Vicuna 70k dataset and Quantised Low Rank Adapations (LoRA).
The Vicuna 70k dataset is a rich collection of user-generated conversations sourced from ShareGPT.com. It encompasses a wide array of topics, including but not limited to:
- Casual conversation
- Storytelling
- Problem-solving
LoRA is a cutting-edge technique designed to reduce the size of large language models without sacrificing performance. It achieves this by:
- Quantizing the Parameters: Reducing the numerical precision of the model's parameters.
- Applying Low-Rank Approximations: Utilizing low-rank approximations on the quantized parameters.
This repository offers a comprehensive guide to fine-tuning the Llama-7B-Uncensored model using the Vicuna 70k dataset with LoRA. It also includes detailed instructions for evaluating the fine-tuned model.
To train the model, simply run the following command:
python train.py configs/open_llama_7b_qlora_uncensored.yaml
Feel free to contribute to this project by submitting issues, pull requests, or reaching out with any questions or suggestions.