POST-TRAINING / DATA INVENTORY
RL Data Atlas.
Find the right tasks for your next run.
Task-curation sources, counts, and reviews in one place.
No matching sources
Try another search or reset the filters.
Counts describe inputs. Counts cover selected input rows at the displayed dataset revision, before conversion or curation. Unknown counts stay blank. Generators have no fixed count; populations may overlap.
Know what is inferred. Families and tags describe the source contract. Benchmark flags reflect declared benchmarks. Difficulty bars link to model attempts and verifier evidence; quality badges link to reviews. Longer bars mean more tasks solved. Difficulty filters and sorting use current Qwen3.5-122B-A10B solve rates only; earlier model identities and budgets are labeled Historical. Bad quality sources are hidden by default.