The small-model storyline this month is "2B beats 4B," and OpenBMB has now published the other half: UltraData-RL-2609, the RL corpus burned during MiniCPM5-2B's post-training, is fully open. The model weights landed on September 7 and the dataset arrived the same day — weights plus data in one drop, still rare among on-device vendors.
The corpus is the L3 refined layer of the UltraData L0-L4 tiered data framework, built for the RL stage: 85,995 tasks, all verifiable-reward — each carries a ground truth a machine can grade, with no free-form LLM judging.
Four directions, every item machine-checkable
The mix is balanced: 32,412 math items (37.7%), 23,665 code (27.5%), 18,046 long-context (21.0%) and 11,872 STEM knowledge (13.8%). Verification differs by domain: math and knowledge answers are matched against reference answers; code submissions are executed against test cases with stdin/stdout compared case by case; long-context items guarantee the context supports the answer. Multiple-choice, true/false, proof, multi-part and image-dependent items were removed from the pipeline outright — what a machine cannot grade does not enter RL.
Six stages, aimed at untrustworthy labels
Stages 4 and 5 carry the weight. Stage 4 audits reward reliability: an LLM judge reviews question-answer consistency, and math/knowledge answers are re-solved by several independent models and relabeled by consensus — items without consensus are dropped, not guessed. Stage 5 calibrates difficulty: repeated rollouts on the RL initialization checkpoint estimate a per-item pass rate; fully mastered items (pass rate 1) are removed, the learnable band is kept, and hard-but-valid items (pass rate 0 with a confirmed label) are retained for online dynamic sampling. This stage changes difficulty and sampling weights only — never the reference label. Upstream sources are all public datasets, reworked: math unions DAPO, DeepScaleR and DeepMath; knowledge draws from OpenScienceReasoning-2; long-context extends HotpotQA, Qasper and MuSiQue with longer contexts plus in-house synthetic data; code builds on the OpenCodeReasoning series plus HardTests.
Measured gains, and three boundaries
The dataset card records that under the JustRL II setting, AIME 2025 rose from 61 to 81 within about 300 RL steps; the final checkpoint that consumed this corpus in post-training reaches 86.5 on AIME 2025 (model card eval table), averaging 53.9 across that table — above the 4B-class models listed for reference (their best is 51.1). The accompanying RL+OPD stage adds an average of 10.96 points on reasoning and general capabilities and 6.96 on agentic tasks.
The boundaries are stated plainly: the code domain ships test cases but no sandbox, so users compute rewards themselves; difficulty filters and online sampling weights are not stored as release fields; decontamination covered only benchmarks known at construction time, and a new check is required before introducing a new benchmark. The license adds a counterintuitive clause: under Apache 2.0, unauthorized unchanged re-hosting, mirroring and commercial repackaging are explicitly prohibited — "open" and "repost freely" are treated as different things.
For teams doing small-model RL, this 188 GB corpus is a rare end-to-end engineering reference: the 2B comeback rests not only on architecture but on the discipline of making every reward trustworthy. The next question is whether the same pipeline can hold this verification density on agentic data.
Refs: UltraData-RL-2609 dataset card / MiniCPM5-2B model card