row_idx int64 | run_id string | problem_id string | sample int64 | template string | effort string | seed int64 | origin string | category string | competition string | source string | nm_uuid string | model string | max_tokens int64 | prompt_tokens_api int64 | completion_tokens_api int64 | reasoning_tokens_api int64 | prompt_len int32 | seq_len int32 | target_len int32 | completion_delta int64 | self_score string | input_ids list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | null | ultra_dsv4_cot_aops_rs0_000000001 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000001 | DeepSeek-V4-Pro | null | null | null | null | 139 | 34,569 | 34,430 | null | null | [
0,
128803,
5718,
1512,
334,
9,
874,
754,
1512,
406,
9,
874,
754,
305,
1512,
345,
9,
27819,
366,
270,
6302,
294,
270,
103609,
294,
260,
2408,
2459,
18908,
1512,
21141,
27819,
13319,
538,
1512,
334,
874,
754,
1512,
406,
874,
754,
305,... |
1 | null | ultra_dsv4_cot_aops_rs0_000000002 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000002 | DeepSeek-V4-Pro | null | null | null | null | 38 | 5,145 | 5,107 | null | null | [
0,
128803,
54204,
270,
16315,
9692,
650,
874,
9884,
16834,
64,
21,
940,
223,
21,
95,
45772,
27707,
76220,
12473,
3986,
3132,
513,
3132,
14,
305,
3065,
782,
4087,
3287,
2549,
874,
7353,
284,
93,
18615,
128804,
128821,
2581,
477,
4869,
... |
2 | null | ultra_dsv4_cot_aops_rs0_000000003 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000003 | DeepSeek-V4-Pro | null | null | null | null | 92 | 32,974 | 32,882 | null | null | [0,128803,25832,270,27773,294,260,5836,83454,13205,49744,260,14770,11938,344,14195,11550,304,260,331(...TRUNCATED) |
3 | null | ultra_dsv4_cot_aops_rs0_000000004 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000004 | DeepSeek-V4-Pro | null | null | null | null | 44 | 1,701 | 1,657 | null | null | [0,128803,34593,270,5793,63895,4015,93,20,2021,4364,14419,19,3069,20,2021,1477,14419,20,2021,929,144(...TRUNCATED) |
4 | null | ultra_dsv4_cot_aops_rs0_000000005 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000005 | DeepSeek-V4-Pro | null | null | null | null | 82 | 4,167 | 4,085 | null | null | [0,128803,22863,1512,67,14,291,14,274,14,283,874,261,874,14229,21451,12149,9487,362,1512,67,2201,295(...TRUNCATED) |
5 | null | ultra_dsv4_cot_aops_rs0_000000006 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000006 | DeepSeek-V4-Pro | null | null | null | null | 51 | 8,008 | 7,957 | null | null | [0,128803,28218,270,18550,6550,957,90,64,20,15138,565,1527,874,4015,93,7828,3069,18437,8119,438,383,(...TRUNCATED) |
6 | null | ultra_dsv4_cot_aops_rs0_000000007 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000007 | DeepSeek-V4-Pro | null | null | null | null | 86 | 28,359 | 28,273 | null | null | [0,128803,5718,957,80,6,366,260,4095,1167,16,1350,1926,260,9069,294,3737,957,1592,1260,288,6,855,786(...TRUNCATED) |
7 | null | ultra_dsv4_cot_aops_rs0_000000009 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000009 | DeepSeek-V4-Pro | null | null | null | null | 54 | 2,737 | 2,683 | null | null | [0,128803,28218,362,1512,1527,874,2605,1512,223,926,874,5982,64,20,1527,874,10915,64,21,1527,438,223(...TRUNCATED) |
8 | null | ultra_dsv4_cot_aops_rs0_000000010 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000010 | DeepSeek-V4-Pro | null | null | null | null | 61 | 27,913 | 27,852 | null | null | [0,128803,671,2499,294,260,1234,48557,1167,305,223,22,344,4939,304,270,2499,294,270,1975,1234,48557,(...TRUNCATED) |
9 | null | ultra_dsv4_cot_aops_rs0_000000011 | null | null | null | 7 | AoPS | aops_cot | null | nemotron-sft-math-v4 | ultra_dsv4_cot_aops_rs0_000000011 | DeepSeek-V4-Pro | null | null | null | null | 62 | 83,607 | 83,545 | null | null | [0,128803,14965,270,9152,3338,1167,294,16005,294,260,40956,7329,70848,295,9692,14229,21451,32594,21,(...TRUNCATED) |
End of preview. Expand in Data Studio
Proof Pilot — training data artifacts
Data-side artifacts for the AIMO Proof Pilot submission, one subdirectory per step of the
Stage 2 "Final Pipeline" report. Code: proof-pilot-codes.
Model checkpoints: proof-pilot-checkpoints.
Each subdirectory contains a SHA256SUMS file.
| Subdirectory | Report step | Contents |
|---|---|---|
step06-teacher-hidden-initial/ |
6 | Extraction manifests for the initial teacher-state corpus: proofs-v2-30k (28,621 docs) and math-v4-aops-cot (51,961 docs), plus the combined 80,582-row index. The multi-TB compressed hidden-state shards themselves are not published; they are regenerated deterministically with the public extraction code and DeepSeek-V4-Flash weights (codec: seed-7 sign/Hadamard rotation + symmetric int6, blocks of 32, fp16 scales) |
step08-proof-agent-65836-snapshot/ |
8–9 | Frozen snapshot of the exact 65,836-turn proof-agent extraction used for the continuation distillation (per_turn.parquet), plus turn_manifest.parquet with the (run_id, problem_id, stage, candidate_id) identity of every turn. The later public materialization with refine/select augmentation (77,538 turns) is dsflash-proof-distill-v2-test |
step09-teacher-hidden-continuation/ |
9 | Extraction manifest (65,836 rows, ~1.39B target tokens), metadata, and the extraction handoff notes for the proof-agent continuation corpus; hidden shards not published (same regeneration path as step06) |
step10-opd-pool-rollouts/ |
10, 12 | Complete OPD dynamic pool (seed + admitted artifacts JSONL) and all admitted rollout parquet shards for both OPD runs (agentic_32b_lc140k, agentic_32b_lc140k_v33), with the full service config.json (admission filter drops finish_reason == "length"). Teacher hidden states were consumed online over HTTP by design and never persisted |
Not included (regenerable from public sources)
- Packed SFT L4 dataset (766,505 × 65,536-token rows, ~50B tokens): regenerate from the public
nemotron-deepseek-sft-mix /
-v2 datasets with the
deterministic builder
training/stage1_v2/build_l4.py(mixmix_g2.json, fingerprintcbe612ceada37eee). - DFlash short/long-context L4 datasets: regenerate with
training/dflash/build_l4_pretokenized.pyfrom the SFT mix (short) and the step10 rollouts + step08 snapshot published here (long). - GPTQ calibration set: sampled deterministically from the SFT L4 bins by
quantization/common.py:load_calib()(fixed seed).
- Downloads last month
- 5