---
language:
- en
license: other
license_name: source-data-and-model-terms
license_link: https://huggingface.co/datasets/WinstonHu/LoC-Path-Data/blob/main/DATA_TERMS.md
tags:
- pathology
- whole-slide-imaging
- visual-question-answering
- conch
- loc-path
configs:
- config_name: stage1_reports
  data_files:
  - split: train
    path: annotations/train/stage1_reports.json
- config_name: stage2_instructions
  data_files:
  - split: train
    path: annotations/train/stage2_instructions.json
---

# LoC-Path Data

> Upload status: complete. All feature files have been checked against the release manifest.

Prepared **WSI-Bench and BCNB** data for LoC-Path training and evaluation:
processed annotation JSON, pre-extracted CONCH-v1 features, and coordinates.
**Raw whole-slide images, image patches, and original pathology reports are not
included.** Model weights are not included.

## Contents

| Component | Published size |
|---|---:|
| Stage 1 report-supervision annotations | 9,635 records |
| Stage 2 instruction annotations | 175,261 records |
| WSI-Bench evaluation annotations (16 task files) | 4,163 records |
| BCNB evaluation annotations | 7,274 records |
| TCGA slides with NPY feature/coordinate pairs | 9,844 |
| BCNB slides with numeric CONCH feature CSVs | 1,058 |
| Numeric feature files | 20,746 |
| Numeric feature download | 340.62 GB (317.23 GiB) |

The WSI-Bench task partitions are inherited from **WSI-LLaVA/WSI-Bench** through
the parent repository's preparation. This release retains those task partitions;
it does not create a new partition. Annotation paths are normalized and records
without available features are filtered. Published counts therefore describe this
prepared snapshot, rather than every upstream annotation version. See
[metadata/dataset_summary.json](metadata/dataset_summary.json) for per-file counts
and [metadata/excluded_records.json](metadata/excluded_records.json) for omissions.

## Layout

```text
annotations/
  train/stage1_reports.json
  train/stage2_instructions.json
  test/wsi_bench/*_test.json
  test/external/bcnb_choice_test.json
features/conch_v1/
  <cancer>_224x224_b20_t15/npy_files/
    <slide>_features.npy
    <slide>_coords.npy
  bcnb_csv/<patient_id>.csv
metadata/
  dataset_summary.json
  feature_manifest.json
  excluded_records.json
  train_test_overlap.json
```

Features have shape `[N, 512]`; coordinates have matching shape `[N, 2]` in `(x, y)`
order. TCGA NPY arrays retain their existing stored precision without a release-time
cast. Model computation follows the training configuration. NPY coordinates use
zero-anchored tile-grid indices. BCNB CSVs contain `patch_name` and 512 numeric
feature columns; patch names encode coordinates, not image contents.

JSON `image` values are **feature lookup stubs**, not downloadable image paths.
For example, `./LUSC/<slide>.h5` resolves to the matching NPY pair in the LoC-Path
training loader. BCNB `feature_path` values are relative to `features/conch_v1`.
Processed report-generation answers are included as supervision; original source
report documents are not distributed.

## Download and use

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="WinstonHu/LoC-Path-Data",
    repo_type="dataset",
    local_dir="data/LoC-Path-Data",
)
```

To download only annotations and documentation initially, pass
`allow_patterns=["annotations/**", "metadata/**", "README.md", "DATA_TERMS.md"]`.
Running the full download again fills in the numeric features.

In the LoC-Path training package:

```bash
export LOCPATH_DATA_ROOT=$PWD/data/LoC-Path-Data
export LOCPATH_FEATURE_ROOT="$LOCPATH_DATA_ROOT/features/conch_v1"
export LOCPATH_STAGE1_REPORT_JSON="$LOCPATH_DATA_ROOT/annotations/train/stage1_reports.json"
export LOCPATH_STAGE1_DATA_PATH="$LOCPATH_STAGE1_REPORT_JSON"
export LOCPATH_STAGE2_DATA_PATH="$LOCPATH_DATA_ROOT/annotations/train/stage2_instructions.json"
```

For WSI-Bench inference, set `DATA_DIR` to
`$LOCPATH_DATA_ROOT/annotations/test/wsi_bench`; for BCNB, use
`$LOCPATH_DATA_ROOT/annotations/test/external`. Set `FEATURE_ROOT` to
`$LOCPATH_FEATURE_ROOT` and pass `--image_feature_suffix .npy` to the inference
launcher. BCNB CSV features are selected from each record's relative path.
No exclusion CSV is needed for this prepared snapshot.

## Sources and terms

Annotations are adapted from [WSI-LLaVA/WSI-Bench](https://github.com/XinhengLyu/WSI-LLaVA)
and [SlideChat](https://github.com/uni-medical/SlideChat). BCNB numeric features
were provided with SlideChat. Features use
[CONCH v1](https://huggingface.co/MahmoodLab/CONCH).
Read [DATA_TERMS.md](DATA_TERMS.md) for attribution and source-data/model terms.
The LoC-Path code license does not relicense these data or the feature extractor.

`metadata/feature_manifest.json` records each numeric file's size, SHA256, feature
shape, stored dtype, and annotation memberships. Local source paths and raw assets
are excluded from the release.