Why this matters
Scaling RL post-training for small-to-mid LLMs demands training signals that are both learnable and checkable. This dataset supplies a curated set of verifiable-reward tasks (≈85,995 samples) designed to give reproducible binary/execute-based feedback rather than noisy heuristic rewards, enabling stable policy improvement in scenarios like MiniCPM post-training and JustRL II experiments.
What Sets It Apart
- Verifiable outcomes by design — every sample encodes a concrete verification mechanism: answer-match for Math/Knowledge/Long-Context and execution-based test cases for Code. This makes rewards unambiguous and audit-friendly for RL algorithms that rely on ground-truth checks.
- Difficulty calibrated for RL — items were filtered and empirically sampled against an RL initialization checkpoint to remove trivially solved items (pass rate 1) while retaining learnable and hard-but-valid items; sampling weights support online dynamic scheduling.
- Multi-domain, RL-oriented formats — 85,995 JSONL samples across Math (32,412), Code (23,665), Long-Context (18,046), and Knowledge (11,872). Each record contains uuid, query, ground_truth, source, and domain, with Code ground_truth carrying paired stdin/stdout arrays for test execution.
- Built for reproducible post-training workflows — used in JustRL II experiments and the MiniCPM5-2B post-training pipeline; construction emphasizes multi-model consensus labeling, LLM-based consistency checks, and cross-validated test cases to improve reward reliability.
Who It's For and Trade-offs
Great fit if you need a compact, high-quality RL corpus to train or evaluate verifiable-reward RL recipes (RLVR/RLPR-style) for LLMs, especially when focusing on math, STEM reasoning, long-context QA, or executable program synthesis. It’s also suitable for mix-ratio studies alongside SFT corpora.
Look elsewhere or prepare extra tooling if you require an integrated execution sandbox (the release contains test cases but no sandbox), need image-dependent or subjective-reward tasks (these were removed), or require dynamic difficulty labels (sampling weights and pass-rate stats used during construction are not included in the release). The project is distributed under Apache-2.0 with upstream licenses retained for derived content.
