OpenBMB · 2026-09-07 · notable
UltraData-RL-2609 — 86,000 checkable RL tasks behind MiniCPM5-2B
UltraData-RL-2609 is an Apache-2.0 reinforcement-learning corpus of 85,995 tasks in math, code, long-context and knowledge. Every item carries a reference answer and a defined way to check it. OpenBMB used it to post-train MiniCPM5-2B.
The open RL corpus OpenBMB used to post-train MiniCPM5-2B — 85,995 tasks that a machine can mark right or wrong.
Key specs
| Parameters | 188 GB |
|---|---|
| Samples | 85,995 |
What is it?
UltraData-RL-2609 collects 85,995 training tasks whose answers can be checked automatically, and OpenBMB published it on Hugging Face on 7 September 2026. The split is 32,412 math problems, 23,665 code tasks, 18,046 long-context questions and 11,872 knowledge questions. It ships as JSONL under Apache-2.0, 188 GB in total, with the upstream sources keeping their own MIT and CC BY / BY-SA 4.0 terms.
How does it work?
Each row holds five fields — uuid, query, ground_truth, source and domain — so a trainer can score an answer without a human or a judge model. Math and knowledge items are marked by answer matching, code by running test cases, and long-context by answer matching against the supplied context. OpenBMB describes six build stages: collection, task standardisation, verifiability filtering, reward-reliability checks, difficulty calibration and quality review. Problems the model already solved every time were thrown out, and hard-but-valid items were kept for dynamic sampling.
Why does it matter?
Reinforcement learning with verifiable rewards only works when the reward signal is trustworthy, and building that check-able task pool is usually the expensive, private part of a post-training run. Publishing UltraData-RL-2609 hands small teams the same corpus that produced MiniCPM5-2B, so a 2B-class model can be pushed on math and code without first assembling and grading tens of thousands of problems. Eight OpenBMB model variants are already listed as trained on it.
Who is it for?
people post-training small open models with RL
Try it
https://huggingface.co/datasets/openbmb/UltraData-RL-2609