314,000 synthetic variational problems generated by the Qwen2.5-32B-Instruct model during 600 steps of RLVR training. Each entry pairs a synthesized mathematical problem with a reference answer generated via the SvS strategy on the DAPO-17k dataset.
Use Cases
- Fine-tune mathematical reasoning models using the synthesized problems and reference answers
- Analyze the diversity of model-generated variational problems across 600 RLVR training steps
- Evaluate the impact of min_hash deduplication at a 0.85 threshold on synthetic mathematical datasets
Strengths
- 314,000 variational problems synthesized by Qwen2.5-32B-Instruct
- Includes reference answers for every synthesized problem instance
- Deduplicated using min_hash with a 0.85 similarity threshold
- Derived from 600 steps of RLVR training using the SvS strategy