STEM tutor
A small open-source model (0.6B parameters) that answers science multiple-choice questions like a tutor.
We built the data (24k STEM multiple-choice questions, preference triplets, 62k Wikipedia and arXiv documents), fine-tuned Qwen3-0.6B with SFT then DPO, added retrieval with a fine-tuned encoder, and quantised the model with QLoRA. The models and datasets are published on Hugging Face.
- accuracy after SFT
- +4 %
- with QLoRA, same accuracy
- 1.1 → 0.5 GB
- +4 % accuracy after SFT, and 73 % reward accuracy for the DPO model.
- QLoRA 4-bit halves the model, from 1.1 to 0.5 GB, at equal accuracy.