Fork of: Stanford Alpaca

This is a fork of the Stanford Alpaca repo with adjustments and additional to enable generation of the in-distribution test dataset and the Sequential Instructions dataset used in Understanding the Effects of RLHF on LLM Generalisation and Diversity. The generated datasets can be found here:

• Sequential Instructions
• In-Distribution Test

To reproduce the generation of the Sequential Instructions dataset, follow the instructions in the Data Generation Process section below, but use python -m generate_instruction_sequential generate_instruction_following_data. This script also has the option of automatically uploading the generated dataset to huggingface using the --save_to_hf=<organisation>/<dataset_name> argument.

For the in-distribution test dataset, follow the instructions in the Data Generation Process section below as-is.

Otherwise, we recommend using the original repository, which has detailed instructions on the rest of the code.

Data Generation Process

Running the code

• Set environment variables OPENAI_API_KEY to your OpenAI API key.
• Install the dependencies with pip install -r requirements.txt.
• Run python -m generate_instruction generate_instruction_following_data to generate the data.
• Optionally pass --save_to_hf=<organisation>/<dataset_name> to automatically upload the generated dataset to huggingface.

Citation

Please cite the original repo if you use the data or code in this repo, as well as our paper:

@misc{alpaca,
  author = {Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto },
  title = {Stanford Alpaca: An Instruction-following LLaMA model},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/tatsu-lab/stanford_alpaca}},
}

@misc{kirkUnderstandingEffectsRLHF2023,
  title = {Understanding the {{Effects}} of {{RLHF}} on {{LLM Generalisation}} and {{Diversity}}},
  author = {Kirk, Robert and Mediratta, Ishita and Nalmpantis, Christoforos and Luketina, Jelena and Hambro, Eric and Grefenstette, Edward and Raileanu, Roberta},
  year = {2023},
  month = oct,
  number = {arXiv:2310.06452},
  eprint = {2310.06452},
  primaryclass = {cs},
  publisher = {{arXiv}},
  doi = {10.48550/arXiv.2310.06452},
  urldate = {2023-10-26},
  archiveprefix = {arxiv},
}

Naturally, you should also cite the original LLaMA paper and the Self-Instruct paper if you use the code or data from this repo.

Acknowledgements

We thank the original Alpaca authors for releasing their code.