Open Models Almanac
Methodsdata curationdata mixture & curriculum

single mention · data curation

RL task sampling proportional to unsolved rate

Dropping always-solved and never-solved tasks, then sampling the remaining tasks in proportion to their unsolved or non-pass rate.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See data mixture & curriculum for the methods filed beside it.

Also called pass-rate buckets.

Filed under data curation › data mixture & curriculum