TamasSimonds/O1-Llama-3.2-3B

O1-Llama 3.2 3B Model Card

Important Disclaimer

This is a proof-of-concept research model designed to demonstrate the feasibility of inducing structured reasoning behaviors in smaller language models. It is not intended for production use or deployment in real-world applications. The model serves primarily as a demonstration of training methodology and should be used only for research purposes.

Model Overview

O1-Llama 3.2 3B is a fine-tuned version of Llama 3.2 3B, trained to demonstrate explicit reasoning patterns similar to those observed in OpenAI's O1 model. The model is trained on ReasonSet, a dataset of worked solutions focusing on mathematical and logical problem-solving.

Key Capabilities

• Explicit brainstorming and strategy enumeration
• Step-by-step solution working out
• Self-correction attempts
• Verification steps in problem-solving

Limitations

• Significantly lower performance compared to larger models
• Can get stuck in circular reasoning
• May fail to find correct solutions despite showing reasoning behavior
• Limited to simpler problems
• Not suitable for production use or critical applications

Training

• Base Model: Llama 3.2 3B
• Dataset: ReasonSet (2,000 worked solutions)
• Domains: AIME, GPQA, MATH dataset problems
• Method: Fine-tuning on worked solutions generated through REL (Reasoning Enhancement Loop)

Intended Use

• Research into reasoning capabilities of smaller language models
• Study of explicit problem-solving behaviors
• Academic investigation of model training methodologies

Repository

Available at: https://github.com/tamassimonds/REL

Citation

[Include paper citation when published]