| # 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] |