The following data files are used for training, testing, and analysis.
| File | Samples |
|---|---|
data/polaris/qwen3-4b-s1-polaris-p1-format.jsonl |
29916 |
data/countdown_3to4/test.jsonl |
1024 |
data/countdown_3to4/train.jsonl |
40000 |
data/countdown_5_ood/test.jsonl |
256 |
data/countdown_5_ood/test_full.jsonl |
1024 |
data/eval/agi_eval_en_prompted.jsonl |
3340 |
data/eval/aime_2024.jsonl |
30 |
data/eval/aime_2025.jsonl |
30 |
data/eval/aime_2026.jsonl |
30 |
data/eval/amc23.jsonl |
40 |
data/eval/arc_c_prompted.jsonl |
1172 |
data/eval/bigbenchhard_prompted.jsonl |
6511 |
data/eval/math500.jsonl |
500 |
data/eval/minervamath.jsonl |
272 |
data/eval/olympic.jsonl |
675 |
data/analysis/polaris_64_seed123.jsonl |
64 |
Countdown: Jiayi-Pan/Countdown-Tasks-3to4 and the existing five-number OOD data; the original parquet files and stats.json are retained as well. The original sources are 3-to-4 Countdown and 5-number Countdown.