3e64138956884942116d3956b0472ef645539db36d9cbc4e5c4a5dd88ef812aa

README.md

@@ -6,70 +6,70 @@ tags:
 - visual-grounding
 - auxiliary-training
 - parent-child
-pretty_name: PathGrounding - ZoomPath Auxiliary Training v4
+pretty_name: PathGrounding - Quality-Filtered Auxiliary Training
 ---
 
 # PathGrounding
 
-Private, portable auxiliary training data for single-step pathology image
-grounding. Each record includes a parent image, an exact child crop, its bounding
-box in parent coordinates, a text question, and organ/MPP/training-level metadata
-when available. There are **618,432 parent-child pairs from seven sources**, using
-**34,193 distinct parent images**.
+Public, portable auxiliary training data for single-step pathology grounding.
+The main package contains **235,346 quality-filtered parent-child pairs from
+seven sources**, with **38,836 distinct parent images**. It replaces the earlier
+unfiltered 618,432-pair main release.
 
-This is not the ZoomPath v4 benchmark, does not include raw WSIs, and does not
-inherit the benchmark's doctor review. Source licenses and restrictions remain
-in the package; RefPath is gated and is not assumed redistributable. Keep this
-repository private unless all upstream redistribution permissions are checked.
+**SegPath is a separate, optional package under `segpath/`: 2,193,058 pairs and
+105,864 parents.** It is not merged into the main manifest, counts, reader or
+sampling configuration. It is disabled by default. Download it explicitly only
+when needed; its mostly small semantic-region boxes differ from the larger
+region-of-interest navigation targets in ZoomPath v4.
 
-## Data
+These are auxiliary training data, not the ZoomPath v4 benchmark. They contain no
+raw WSIs and have not inherited the benchmark's doctor/pathologist review.
+Upstream licenses and restrictions apply; RefPath is gated and redistribution
+permission is not assumed; users must follow all relevant upstream terms.
+SegPath is CC-BY-NC-SA 4.0.
 
-| Source | Pairs |
+## Main Package
+
+| Source | Retained pairs |
 | --- | ---: |
-| RefPath train | 28,908 |
-| PanNuke | 187,831 |
-| NuInsSeg | 25,718 |
-| NuCLS | 35,848 |
-| BCSS | 6,153 |
-| Lizard | 325,966 |
+| BCSS | 2,836 |
+| Lizard | 31,446 |
 | MIDOGpp | 8,008 |
-| **Total** | **618,432** |
-
-`partition` is `training`. Original `source_split`/fold information is preserved;
-no new train/val/test split is imposed. Keep targets from the same parent together
-when defining a later split. Nominal MPP and training scale estimates are flagged
-as approximate, especially the heuristic 20x/40x assignment for RefPath.
+| NuCLS | 32,538 |
+| NuInsSeg | 20,521 |
+| PanNuke | 111,605 |
+| RefPath train | 28,392 |
+| **Main total, excluding SegPath** | **235,346** |
 
-## Download And Restore
+Each record includes parent/child images, an exact parent-relative child box,
+text question, organ and available scale metadata. `partition` is `training`;
+original source split/fold information is retained, with no new train/val/test
+split. Keep all related targets/parents, and source slides when identifiable,
+together when defining later splits.
 
-Python 3.10+ is required. Authenticate with an account that has repository access:
+Python 3.10+ is required. Authenticate with an account that has repository access.
+The explicit allow-list below does **not** download SegPath:
 
 ```bash
 pip install -U huggingface_hub pillow
 hf auth login
-hf download xuxuxuxuxu/PathGrounding \
-  --repo-type dataset --local-dir /data/pathgrounding-download
-python /data/pathgrounding-download/tools/restore_package.py \
-  /data/pathgrounding-download /data/zoompath_training_v4
 ```
 
-Alternatively, call `huggingface_hub.snapshot_download` with the same repository,
-`repo_type="dataset"`, and `local_dir="/data/pathgrounding-download"`, then run
-the restore command. Interrupted downloads can be resumed.
-
-The images and per-parent annotations are stored in uncompressed TAR shards
-(approximately 2 GiB each) rather than 686,000 individual repository files.
-Restoration verifies every shard's SHA-256 checksum and recreates the original
-relative paths. It does not resize images or rewrite annotations/questions.
-Restoration can be rerun after interruption; it rewrites image/annotation files
-from the verified archives. Use a dedicated output directory.
-
-The data package is approximately 54 GB. Allow roughly 115 GB of free space to
-retain both downloaded archives and the restored package.
-
-## Training Reader
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id="xuxuxuxuxu/PathGrounding",
+    repo_type="dataset",
+    local_dir="/data/pathgrounding-download",
+    allow_patterns=["README.md", "release_manifest.json", "package/**",
+                    "archives/**", "tools/restore_package.py"],
+)
+```
 
 ```bash
+python /data/pathgrounding-download/tools/restore_package.py \
+  /data/pathgrounding-download /data/zoompath_training_v4
 export PYTHONPATH=/data/zoompath_training_v4:$PYTHONPATH
 ```
 
@@ -79,37 +79,89 @@ from training_data import AuxiliaryTrainingReader
 reader = AuxiliaryTrainingReader("/data/zoompath_training_v4")
 row = next(reader.iter_records(dataset="refpath_train"))
 sample = reader.load_pair(row)
-
-parent = sample["parent_image"]
-child = sample["child_image"]
+parent, child = sample["parent_image"], sample["child_image"]
 question = sample["question"]
 selected_box = sample["child_bbox_xyxy_in_parent"]
-all_valid_boxes = sample["valid_target_bboxes_xyxy_in_parent"]
+valid_boxes = sample["valid_target_bboxes_xyxy_in_parent"]
+ignored_boxes = sample["ignored_target_bboxes_xyxy_in_parent"]
 organ, mpp, level = sample["organ"], sample["mpp"], sample["level"]
 ```
 
-Coordinates are integer pixel XYXY, top-left origin, half-open edges. The child
-is exactly the selected parent crop. `level` is `high`/`mid`/`low`, not a WSI
-pyramid index; MPP is in microns per pixel. Unsupported metadata remains null.
+The main images and annotations are uncompressed TAR shards, approximately
+2 GiB each. Restoration verifies SHA-256 and recreates relative paths without
+resizing or changing questions. It is resumable by rerunning the command.
+Use a dedicated output directory and allow space for both the downloaded shards
+and the restored package. Exact byte totals are in `release_manifest.json`.
+
+Coordinates are integer pixel XYXY, top-left origin, half-open edges. Children
+are exact crops. For multi-target supervision, use all eligible `valid_boxes`,
+not every annotated box or only the selected one. Source-supported answers
+rejected by quality filtering remain `ignored_boxes`, never negatives.
+The reader checks the exact label/expression and target count. Stream JSONL;
+uniform row sampling is not uniform parent/query sampling.
+
+`level` means the training category `high`/`mid`/`low`, not a WSI pyramid index.
+MPP is microns per pixel; nominal RefPath magnification/MPP are approximate and
+explicitly flagged. Unsupported metadata remains null.
+
+## Optional SegPath
+
+Download only the independent SegPath package:
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id="xuxuxuxuxu/PathGrounding",
+    repo_type="dataset",
+    local_dir="/data/pathgrounding-download",
+    allow_patterns=["segpath/**", "tools/verify_segpath_package.py"],
+)
+```
+
+```bash
+python /data/pathgrounding-download/tools/verify_segpath_package.py \
+  /data/pathgrounding-download/segpath
+export PYTHONPATH=/data/pathgrounding-download/segpath:$PYTHONPATH
+```
+
+Use `AuxiliaryTrainingReader("/data/pathgrounding-download/segpath")` in a
+separate reader/import context from the main package. The SegPath reader loads
+images and annotations directly from eight native, uncompressed marker TARs;
+no extraction is required. Reserve approximately 210 GiB plus download-cache
+headroom. `segpath/release_manifest.json` gives the exact total and checksums.
+
+Only official training patches are included. Targets are 8-connected semantic
+mask regions, **not cell instances or WSI navigation paths**. Questions use full
+labels. Known organs are retained; unknown organs are null. All rows record 40x,
+approximate MPP `[0.25, 0.25]`, and training level `high`.
 
-For multi-target training, use **all valid boxes**, not every box in the parent
-annotation file and not just the selected box. The reader applies the exact
-label/expression filter and checks `target_count`. Unselected matching targets
-are valid alternatives, not negatives. Stream `manifest.jsonl`; row-uniform
-sampling is not parent/query-uniform sampling.
+SegPath's `sampling_config.json` sets `enabled_by_default: false` and
+`default_mix_probability: 0.0`. No combined manifest is provided. An explicitly
+chosen future experiment should cap its contribution and sample marker, then
+parent/query, then target, rather than concatenate rows and sample uniformly.
 
 ## Layout And Validation
 
 | Repository path | Contents |
 | --- | --- |
-| `package/` | Manifest, index, reader, documentation, validation and source metadata |
-| `archives/` | Parent images, child crops and all per-parent annotations |
-| `tools/restore_package.py` | Checksum verifier and portable restoration tool |
-| `release_manifest.json` | File checksums, sizes, archive inventories and source counts |
-
-The restored package's `README.md` documents the full schema, provenance,
-approximate scale mapping, valid-target semantics and source restrictions.
-`validation.json` reports the completed export audit: all rows checked for
-structure/metadata and 700 deterministic crop pixel samples, with zero errors.
-The original source audit is preserved separately; the 700-sample export audit
-is not a claim that every exported image was rechecked pixel-by-pixel.
+| `package/` | Main manifest, index, reader, quality policy, decisions, reports and provenance |
+| `archives/` | Main parent/child images and per-parent annotations |
+| `release_manifest.json` | Main-only inventory, counts, sizes and SHA-256 checksums |
+| `tools/restore_package.py` | Main package checksum verification and restoration |
+| `segpath/` | Independent optional SegPath manifest, index, TARs, reader and provenance |
+| `segpath/release_manifest.json` | SegPath-only inventory, counts, sizes and checksums |
+| `tools/verify_segpath_package.py` | SegPath checksums and sample loading check |
+
+Both packages report `pass` with zero errors in their own `validation.json` and
+`quality_validation.json`. The main export checks all 235,346 rows and 38,836
+annotation views, with 700 deterministic crop-pixel samples. SegPath checks all
+2,193,058 rows and 105,864 annotation views, with 1,000 deterministic crop-pixel
+rechecks; its source audit reconstructs every retained parent's original mask.
+These automated audits are not a claim of a new manual or physician review.
+
+The main package preserves historical source metadata showing SegPath excluded
+from that build; this is intentional and does not invalidate the independent
+validated `segpath/` release. Read each package's methods, quality policy,
+filter results and provenance before training. Pin a Hub commit revision for
+reproducible downloads.

archives/annotations-00000.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:88cd521c4fb64ca027fea2f348bba39a314a55dad91b9a028e52361cdf243bd3
-size 285358080
+oid sha256:31f467ba323acf29cface50a8b4debcee8581d6dd8a4e3a7eb53e56e3533410d
+size 621434880