# Processed 40k candidate datasets (repaired v3)

These are the historical **OpenMathInstruct-2-derived** datasets, not the new
deduplicated OpenMathReasoning trace collection. They were downloaded from NRP
(`nautilus`, namespace `wcsng-log`, PVC `reasoning-test`) on 2026-09-11.
Source files live under `/data/datasets/` with the same basenames as the local files.
The remote files were not modified. Superseded v1/v2 repairs are not included.

Each `openmathinstruct2_plans_40k_candidateN_repaired_v3.jsonl` contains 40,000
records, including the original generation fields and the derived `repaired_output`
field. Each dataset has `.summary.json`, `.validation.json`, `.audit.json`, and
`.manual_review.jsonl` sidecars. No rows have been filtered from these downloads.

| Candidate | Format | Historically strict-valid rows | Unrecoverable rows |
|---|---|---:|---:|
| 1 | One-line plan | 39,999 | 1 |
| 2 | Summarized working | 39,995 | 5 |
| 3 | Action-state trajectory | 39,994 | 6 |

Structural validity and a successful generation request do not establish
mathematical correctness or semantic fidelity. Review the repair and validation
sidecars before selecting training rows.

## Frozen content identities

All three local dataset SHA-256 hashes and byte sizes were verified against the
repository's historical freeze manifest after download on 2026-09-11. Each local
dataset also has exactly 40,000 lines. The compressed transfer completed normally
and included all 12 sidecars; structural validation was not rerun during download.

| Candidate | Bytes | SHA-256 |
|---|---:|---|
| 1 | 1316722574 | `2043422dd55aa1b506a0ed53bbae232218cf1850d652c2014a7c4fc041da5fbb` |
| 2 | 1387069920 | `152e84f999ed271fa6d54e6aadecdac65d47dbcaef1e99889f9c7a46c65b4e5b` |
| 3 | 1430324162 | `86013a9548a987d964fc1074c570f2c5f5577592bcf1b5b8a4c69d2f74110898` |

Reference: [historical freeze manifest](../../openmathreasoning/reports/openmath_target_freeze_v1_20260829.json)
and [repair report](../../openmathreasoning/reports/openmath_repair_v3_20260828.md).
Those historical reports use OpenMathReasoning terminology in places; the actual
dataset provenance here is OpenMathInstruct-2.

The downloaded JSON/JSONL artifacts are ignored by the directory's `.gitignore`.