Datasets:
Frozen task examples
Some third-party task materials are omitted from this release pending redistribution permission. See the coverage manifest for included tasks.
The JSONL files contain actual inputs, targets, and complete prompts from archived evaluation runs. Selection prefers a main run, then OLMo-2 13B, then the other models in the release inventory. This is one reference snapshot per task, not all prompt variants across every checkpoint. No instances were regenerated.
Use prompt verbatim, including its demonstrations, to replay that reference input. target is the saved scoring target; input is the saved query when available. Record IDs distinguish rows; content-based example IDs can repeat for duplicate examples. Metadata records run settings where available and differences from the appendix's example counts. These differences are preserved, not silently repaired.
sources/ contains the current task/scoring implementations for reference. They are not a standalone evaluation package, and their historical revision is unknown. Loading the frozen records needs only Python's standard library.
IOI material retains its upstream MIT license; see THIRD_PARTY_NOTICES.md. Other included project task material is CC BY 4.0.