Model Overview
allenai/open-instruct-cot-7b is a 7 billion parameter LLaMa model developed by AllenAI, specifically fine-tuned on the Chain-of-Thought (CoT) dataset, which is a subset of Flan v2. This model was created as part of the research presented in the paper "How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources." It is distributed as a model diff, meaning users need to apply it to an existing LLaMa base model to recover the full fine-tuned model.
Key Capabilities
- Instruction Following: Optimized for understanding and executing instructions, particularly those benefiting from Chain-of-Thought reasoning.
- Reasoning Tasks: Fine-tuning on the CoT dataset aims to improve performance on complex reasoning benchmarks.
- Benchmark Performance: Achieves a 22.4% average across various benchmarks including MMLU, GSM, BBH, TydiQA, and Codex-Eval, with specific scores like 27.5% on GSM CoT and 31.3% on BBH CoT.
Usage and Input Format
To use this model, users must first have access to a LLaMa model in Hugging Face format. The allenai/open-instruct-cot-7b diff is then applied using a provided script to recover the full model. The model expects inputs formatted with specific tokens:
<|user|>
Your message here!
<|assistant|>
It is crucial to include a newline after <|assistant|> for optimal generation quality. The codebase for training and evaluation is available on GitHub.