Finetuned Qwen1.5B (R1 Distilled Version) on this dataset, which comes from this dataset but with an additional "summary" produced by an in-house synthetic data generator.

This LoRA is therefore a LoRA which helps the model return a "\n\nFinal Answer: ..." after it's reasoning and initial response steps.

See this paper for more details.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.