FineProofs RL: 5,227 Olympiad Math Problems with RL Rubrics

5,227 mathematical Olympiad problems and rubrics for reinforcement learning training, released by lm-provers in 2026. The collection includes problems from international competitions and Art of Problem Solving (AoPS) annotated with Gemini-3-Pro rubrics and Qwen-based rewards.

Use Cases

• Training reward models using the 0-7 point rubrics
• Fine-tuning LLMs for mathematical reasoning using per-rollout scores
• Benchmarking AI performance on international Olympiad-level problems

Strengths

• 5,227 problems from international Olympiads and AoPS
• Rubrics use a specific 0-7 point scale
• Includes per-rollout scores from Qwen/Qwen3-4B-Thinking-2507

Limitations

• Rubrics and scores are synthetic/model-generated rather than human-verified
• Limited to 5,227 records, which may be small for some large-scale RL tasks

Provenance

Source

International Olympiad competitions and Art of Problem Solving (AoPS)

Collection Method

Sourced from competitions and AoPS, with synthetic annotations generated by Gemini-3-Pro and Qwen models

Freshness

Last updated February 2026.

Geography

Global