This repository provides the official code for "Large Language Models Are Reasoning Teachers," an ACL 2023 paper. It enables users to run Chain-of-Thought (CoT) reasoning on OpenAI models and fine-tune student models (like T5, Flan-T5, GPT-2) using the Fine-tune-CoT method on custom datasets and hardware.

How It Works

The project implements two main functionalities: OpenAI API experiments and custom experiments on user-provided hardware. For OpenAI, it leverages their API for data collection and fine-tuning. For custom experiments, it utilizes PyTorch Lightning and Hugging Face Transformers to fine-tune open-source models. The core methodology, Fine-tune-CoT, trains student models using reasoning data generated by larger "teacher" models, aiming to transfer complex reasoning capabilities.

Quick Start & Requirements

• Install: pip install -r requirements.txt and python setup.py develop
• Environment: Python <= 3.10, PyTorch Lightning <= 1.9, PyTorch >= 2.0
• Data: Download experimental data from provided Dropbox/Google Drive links (approx. 8GB for completion data).
• Notebooks for OpenAI API experiments and custom fine-tuning are available.

Highlighted Details

• Code for running CoT reasoning on OpenAI models.
• Fine-tune-CoT implementation for training student models (T5, Flan-T5, GPT-2) on custom GPUs.
• All experimental data is publicly shared in JSON/JSONL formats.
• Includes scripts to reproduce paper results and generate tables/figures.

Maintenance & Community

• Developed by Namgyu Ho, Laura Schmid, and Se-young Yun.
• Accepted to ACL 2023.
• No explicit community links (Discord/Slack) or roadmap are provided in the README.

Licensing & Compatibility

• The repository is released under the MIT License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The project has specific Python and PyTorch version requirements. While it supports custom open-source models, the primary focus and extensive data are geared towards OpenAI models. The README mentions "Needs update" for data organization patterns, suggesting potential ongoing development or changes.