| { | |
| "_doc": "A decoder training corpus built from the drawn split. The bar is the CAMPAIGN's: every part here had its program judged against its exact CAD surfaces and passed. This exporter re-tokenizes; it does not re-verify. `rejected` counts records that tokenized but could not be trained on, by reason -- a corpus that shrinks silently is how a run ends up measuring something else.", | |
| "split_rev": "abc-v1-provisional", | |
| "bar": "hd-exact-v3", | |
| "generated_at": "2026-07-28T05:24:00.172406+00:00", | |
| "dense_nbins": 256, | |
| "candidates": 136719, | |
| "skipped_before_export": {}, | |
| "written": 133238, | |
| "by_split": { | |
| "dev": 5136, | |
| "holdout": 10001, | |
| "train": 118101 | |
| }, | |
| "rejected": { | |
| "target_did_not_roundtrip": 3481 | |
| }, | |
| "shard_failures": [], | |
| "pairs": "/mnt/qn45/hexforge-data/corpora/hexgen_abc_v1/pairs.jsonl" | |
| } | |