DELVE
Collection
Discovery Execution Framework for Mathematical Reasoning • 4 items • Updated • 1
Dataset Viewer
The dataset viewer is not available for this split.
Job has been terminated due to a temporary spike in resource usage and may be restarted later.
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
DELVE model outputs
Native model trajectories and correctness audits from DELVE. Each compute block (1×) allows 8,192 output tokens; 5× allows 40,960.
Files follow <competition>/<year>/<model>/<arm>/<problem_id>/seed_<n>/. Each arm also contains audit.jsonl.
| Arm | Meaning |
|---|---|
unaided-1x |
Independent one-block runs; AIME has 32 seeds (0–31) per problem. |
unaided-4x |
DeepSeek unaided baselines through 4×, with four seeds (0–3). |
unaided-8x |
Earlier DeepSeek AIME unaided baselines through 8×. |
unaided-5x |
Unaided runs through 5×, with eight seeds (0–7). |
oracle-at-0x-execution |
A frozen solution sketch injected before generation; eight seeds through 5×. |
continue-at-kx |
Eight unaided branches from checkpoint k× of unaided source seed k−1, for k=1–4. |
continue-oracle-at-kx |
The same branches, with the frozen sketch injected at the fork. |
Branches preserve the source conversation and generate only through the absolute 5× budget. Child seed numbers identify the eight branches, separately from the source seed. HMMT and BRUMO have no independent unaided-1x arm in this experiment series.
Each trajectory contains meta.json, compressed checkpoint_<k>x.json.zst and conversation.json.zst, and audit.json. An audit records checkpoint correctness; cumulative solved counts include success at any earlier checkpoint, including an inherited prefix. Only completed, verified batches are published.
- Downloads last month
- 146