king159 commited on
Commit
2acb544
·
verified ·
1 Parent(s): 53676f4

Align the GPSBench dataset card with the release documentation

Browse files
Files changed (2) hide show
  1. README.md +39 -11
  2. checksums.sha256 +1 -1
README.md CHANGED
@@ -3,13 +3,8 @@ pretty_name: GPSBench
3
  license: other
4
  license_name: dataset-source-terms
5
  license_link: LICENSE
6
- language:
7
- - en
8
- tags:
9
- - spatial-transcriptomics
10
- - rna-localization
11
- - point-cloud
12
- - biology
13
  configs:
14
  - config_name: NIH3T3
15
  data_files:
@@ -36,9 +31,29 @@ configs:
36
  - split: test
37
  path: data/Lung_test/Lung_test_cells.parquet
38
  ---
39
- # GPSBench
40
 
41
- GPSBench evaluates multi-label subcellular RNA localization from individual cells’ molecular point clouds. Its three platform-specific tasks use separate vocabularies and label sets. Each task keeps the field-of-view split of the processed input snapshot used for GPSNet. The [GPSNet repository](https://github.com/king159/GPSNet) contains the model, preprocessing, and evaluation code.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
 
43
  ## Sources and reference labels
44
 
@@ -76,12 +91,25 @@ The benchmark inputs derive from the fixed [st-data-distil snapshot](https://hug
76
  - `staging/Lung/fov_<fov>.parquet` contains the filtered Lung molecules with their cell IDs, gene symbols, global and local pixels, `z`, and compartment flags. The GPSNet materializer reconstructs the per-cell Parquet layout and computes the normalized coordinates and 256-neighbor cache. Lung boundary shapes originate from the split annotations in FOV-local pixels, and some are null.
77
  - The root `manifest.json` identifies the input revision, included splits, and archive files. `checksums.sha256` verifies the files in a downloaded snapshot.
78
 
79
- You can inspect cell metadata with `datasets.load_dataset("king159/GPSBench", "NIH3T3", split="train")`. For the full model inputs, download into a fresh local directory, verify the snapshot, and unpack the spot archives on Linux:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80
 
81
  ```bash
82
  (
83
  set -e
84
- hf download king159/GPSBench --repo-type dataset --local-dir gpsbench
85
  cd gpsbench
86
  sha256sum --check checksums.sha256
87
  if [ -d spots ]; then
 
3
  license: other
4
  license_name: dataset-source-terms
5
  license_link: LICENSE
6
+ language: [en]
7
+ tags: [biology, spatial-transcriptomics, rna-localization, point-cloud]
 
 
 
 
 
8
  configs:
9
  - config_name: NIH3T3
10
  data_files:
 
31
  - split: test
32
  path: data/Lung_test/Lung_test_cells.parquet
33
  ---
34
+ <h1 align="center">GPSBench: Subcellular Gene Spatial Pattern Classification</h1>
35
 
36
+ <p align="center">GPSBench evaluates gene-level RNA localization from molecular point clouds across seqFISH+, MERFISH, and CosMx imaging data.</p>
37
+
38
+ <p align="center">
39
+ <a href="https://personal-page.wjh.app/GPSNet/">Project page</a> ·
40
+ <a href="https://personal-page.wjh.app/GPSNet/files/gpsnet-paper.pdf">Paper</a> ·
41
+ <a href="https://github.com/king159/GPSNet">Code</a>
42
+ </p>
43
+
44
+ ## Dataset overview
45
+
46
+ This release contains **3 datasets, 101,347 cells, and 43,663,730 RNA molecules**.
47
+
48
+ | Dataset | Cells | RNA molecules | Genes |
49
+ | --- | ---: | ---: | ---: |
50
+ | NIH3T3 | 179 | 2,724,808 | 3,721 |
51
+ | U2OS | 1,022 | 10,634,467 | 130 |
52
+ | Lung | 100,146 | 30,304,455 | 960 |
53
+
54
+ Each task has a separate gene vocabulary and label set. The train, validation,
55
+ and test splits follow disjoint fields of view and are shared across methods
56
+ and seeds.
57
 
58
  ## Sources and reference labels
59
 
 
91
  - `staging/Lung/fov_<fov>.parquet` contains the filtered Lung molecules with their cell IDs, gene symbols, global and local pixels, `z`, and compartment flags. The GPSNet materializer reconstructs the per-cell Parquet layout and computes the normalized coordinates and 256-neighbor cache. Lung boundary shapes originate from the split annotations in FOV-local pixels, and some are null.
92
  - The root `manifest.json` identifies the input revision, included splits, and archive files. `checksums.sha256` verifies the files in a downloaded snapshot.
93
 
94
+ With Hugging Face Datasets installed, inspect cell metadata with
95
+ `datasets.load_dataset("king159/GPSBench", "NIH3T3", split="train")`.
96
+ Download the full bundle into a fresh directory, then verify the snapshot and
97
+ unpack the spot archives on Linux.
98
+
99
+ ```python
100
+ from huggingface_hub import snapshot_download
101
+
102
+ snapshot_download(
103
+ repo_id="king159/GPSBench",
104
+ repo_type="dataset",
105
+ local_dir="gpsbench",
106
+ ignore_patterns=[".gitattributes"],
107
+ )
108
+ ```
109
 
110
  ```bash
111
  (
112
  set -e
 
113
  cd gpsbench
114
  sha256sum --check checksums.sha256
115
  if [ -d spots ]; then
checksums.sha256 CHANGED
@@ -1,6 +1,6 @@
1
  50e88a842e5f5ad786a4f9825e32a9d8d268f08af2f7c890ad23beb020c15630 Bento-data-MIT.txt
2
  0b236ef789b2da529e6d85df25e2a01b0efae8e16361c610da31221e452bd21b LICENSE
3
- 23cbff286371c31d70e4782df0b49d575401fbd47bc4df75ffeb0fd2f165c91a README.md
4
  f1bf2c635495a40d17bf54f81b7ab89fafc3306bbe120f1829b2c0566823ca67 data/Lung_test/Lung_test_cells.parquet
5
  ada005b37a48b1c3d2910655aa943612a17069ac63c25b6dc653f2994bac249d data/Lung_test/manifest.json
6
  fffcea9cb2bcec38c8e302e11700084b5c071d814c07f20a5df26153099cf384 data/Lung_train/Lung_train_cells.parquet
 
1
  50e88a842e5f5ad786a4f9825e32a9d8d268f08af2f7c890ad23beb020c15630 Bento-data-MIT.txt
2
  0b236ef789b2da529e6d85df25e2a01b0efae8e16361c610da31221e452bd21b LICENSE
3
+ e307e4300f671ebfef33e267641312b377684c90993535309dc6c5054a74cde1 README.md
4
  f1bf2c635495a40d17bf54f81b7ab89fafc3306bbe120f1829b2c0566823ca67 data/Lung_test/Lung_test_cells.parquet
5
  ada005b37a48b1c3d2910655aa943612a17069ac63c25b6dc653f2994bac249d data/Lung_test/manifest.json
6
  fffcea9cb2bcec38c8e302e11700084b5c071d814c07f20a5df26153099cf384 data/Lung_train/Lung_train_cells.parquet