| language: en | |
| tags: [autoscientist, adaption-labs, qlora, mathcode, dpo] | |
| base_model: Qwen/Qwen2.5-0.5B-Instruct | |
| datasets: [Rishidar/autoscientist-mathcode-dataset] | |
| # AutoScientist Competition — Mathcode Model | |
| Qwen2.5-0.5B-Instruct adapted for **mathcode** via Adaption Labs AutoScientist v5: | |
| 4-bit QLoRA SFT (r=32, alpha=64) then DPO (beta=0.1) on chosen/rejected pairs. | |
| - DPO reward accuracy: 0.8181818181818182 | |
| - DPO reward margin: 8.761996030807495 | |
| Dataset: [Rishidar/autoscientist-mathcode-dataset](https://huggingface.co/datasets/Rishidar/autoscientist-mathcode-dataset). | |
| Also mirrored on Kaggle: rishidard/autoscientist-mathcode-qlora. | |