This repository provides the "CoT Collection," a dataset of 1.84 million Chain-of-Thought rationales across 1,060 tasks, aimed at improving zero-shot and few-shot learning in language models. It also offers pre-trained CoT-T5 models (11B and 3B parameters) leveraging this fine-tuning approach. The project targets researchers and practitioners seeking to enhance LLM reasoning and generalization, especially in data-scarce scenarios.

How It Works

The core innovation is fine-tuning language models using extensive Chain-of-Thought (CoT) rationales. This approach imbues models with explicit reasoning steps, improving generalization and performance on unseen tasks with minimal examples. Scripts are provided for augmenting datasets with CoT rationales, typically generated via external LLM APIs like OpenAI's, which are then used to train specialized models such as CoT-T5. This method directly teaches models step-by-step thinking, leading to more robust predictions.

Quick Start & Requirements

Access the CoT Collection dataset via Hugging Face's datasets library: from datasets import load_dataset; dataset = load_dataset("kaist-ai/CoT-Collection"). Pre-trained CoT-T5 models (11B and 3B versions) are available via Hugging Face's transformers: from transformers import AutoTokenizer, AutoModelForSeq2SeqLM; tokenizer = AutoTokenizer.from_pretrained("kaist-ai/CoT-T5-11B"); model = AutoModelForSeq2SeqLM.from_pretrained("kaist-ai/CoT-T5-11B"). Rationale augmentation scripts require OpenAI API keys. Paper: https://arxiv.org/abs/2305.14045.

Highlighted Details

• Dataset Size: 1.84 million Chain-of-Thought rationales.
• Task Coverage: 1,060 diverse tasks.
• Model Availability: CoT-T5 models in 11B and 3B parameter sizes.
• Methodology: Fine-tuning via explicit reasoning chains enhances LLM generalization.

Maintenance & Community

Project code is currently refactoring and will be uploaded soon. Contact seungone@kaist.ac.kr for inquiries. No community channels are explicitly mentioned.

Licensing & Compatibility

Dataset and models are strictly for non-commercial use. Usage is subject to OpenAI's Terms of Use for generated data. This restricts application in commercial products.

Limitations & Caveats

The primary limitation is the non-commercial license, precluding business use. Project code is not yet available due to refactoring. Rationale augmentation depends on external API keys (OpenAI), introducing dependency and potential costs.