| # SegPath Auxiliary Training v1 |
|
|
| ## Location |
|
|
| SegPath is processed separately from the frozen v4 benchmark and from the seven |
| already-filtered auxiliary sources. The destination is: |
|
|
| `/jhcnas3/jiahao/ZoomPath/auxiliary_training_v2/segpath_training_sharded` |
|
|
| The processing inputs are the previously acquired official eight-marker source |
| release at `/home/jiahao/jiahao/ZoomPath/auxiliary_training_v2/sources/segpath`. |
| That input path resolves to `/storage/ssd1/jiahao/ZoomPath/auxiliary_training_v2/sources/segpath`. |
| Inputs are read in place; large new outputs are written to NAS, not SSD. |
| The original compressed source archives are separately copied to NAS. |
| Do not remove SSD originals until that copy has finished and been verified. |
|
|
| ## Construction |
|
|
| - Use official CSV `train` patches only; exclude `val` and `test`. |
| - Derive bounding boxes from original binary-mask 8-connected components. |
| - Store semantic regions, not cell-instance or WSI-navigation ground truth. |
| - Keep same-label eligible targets as alternative answers in shared annotations. |
| - Mark quality-rejected supported targets as ignored, never negatives. |
| - Use full-label questions: `Find a region that shows {label}.` |
| - Preserve original parent pixels and exact child crops without resizing. |
| - Record 40x, nominal approximate MPP 0.25 and training level `high`. |
| - Map organ conservatively from the paper's Table S2; uncertain sites are null. |
| - Store parent/child PNGs and JSON annotations in eight uncompressed TAR shards. |
| - Keep a plain JSONL manifest with package-relative TAR member byte ranges. |
|
|
| Detailed thresholds and limitations live in |
| `scripts/segpath_quality_policy_v1.json` and |
| `scripts/segpath_training_METHODS.md`; both are copied into the training package. |
| The package reader requires no extraction and works after relocating the whole |
| package. Its documentation includes loading and sampling examples. |
|
|
| ## Checks and Release |
|
|
| Construction pixel-compares every emitted child. Then run the portable export |
| validator over all rows and a deterministic pixel sample. Independently rebuild |
| every retained parent's semantic components from its original mask and compare |
| the official split, target identities, decisions and counts. |
|
|
| The definitive counts and state are the generated `index.json`, |
| `quality_filter_report.json`, `validation.json` and `quality_validation.json`. |
| Only a package passing both validators is ready for training. No new physician |
| review is claimed, and no v4 benchmark files or Hugging Face release are changed |
| by this processing step. |
|
|
| SegPath is optional and disabled by default, with mixture probability 0.0. Do not |
| merge its manifest or sampling with the seven-source main package. If a later |
| experiment explicitly enables it, cap its contribution and sample marker, then |
| parent/query, then retained target. Raw row-uniform sampling would overrepresent |
| dense parents. The upstream CC-BY-NC-SA 4.0 restrictions still apply. |
|
|
| ## Final Audit on 2026-10-10 |
|
|
| - Official source patches: 158,687 across eight markers. |
| - Official training parents checked: 133,358; heldout patches excluded: 25,329. |
| - Candidate source components >=128 mask pixels: 2,819,546. |
| - Retained training pairs: 2,193,058 across 105,864 unique parents. |
| - Quality-rejected components: 626,488; 619,421 remain ignored in retained views. |
| - Every emitted child passed exact parent-crop pixel comparison during build. |
| - Portable validator: all 2,193,058 rows, 105,864 annotations, global ID/count |
| checks and 1,000 deterministic crop rechecks; pass with zero errors. |
| - Source validator: all 105,864 retained parents reconstructed from original |
| masks and checked against official CSVs; pass with zero errors. |
| - Training image/annotation TAR payload storage: 215,130,941,440 bytes |
| (approximately 200.36 GiB), plus JSONL manifests, checkpoints and audit data. |
| - Eight raw compressed archives were copied to NAS and all SHA-256 hashes match |
| the downloaded release, total 305,525,589,629 bytes (approximately 284.54 GiB). |
| - SSD source originals were not deleted. Frozen benchmark and Hugging Face |
| repositories were not changed. |
|
|
| The canonical reports are inside the NAS package. Source-copy verification is |
| `/jhcnas3/jiahao/ZoomPath/auxiliary_training_v2/sources/segpath/nas_archive_verification.json`. |
|
|
