Tighten README: trim noise, fix repo names, ASCII-only.

README.md

@@ -8,9 +8,7 @@ task_categories:
 tags:
 - cell-painting
 - microscopy
-- representation-learning
 - single-cell
-- biology
 - jump-cp
 configs:
 - config_name: default
@@ -19,138 +17,90 @@ configs:
     path: data/shard-*.tar
 ---
 
-# CP-BG-Bench: JUMP view SD (background-masked + density-augmented cell crop)
+# CP-BG-Bench JUMP SD: background-masked + density-augmented cell crop
 
-This dataset is one of four paired views of the same single-cell tiles from
-the JUMP-CP Cell Painting consortium, prepared for the CP-BG-Bench evaluation
-framework.
+Single-cell tiles from JUMP-CP, packaged for the CP-BG-Bench evaluation.
+This is one of four row-aligned views of the same 645,390 cells:
 
-The four views — C (raw crop), S (background-masked crop), CD (density-augmented),
-SD (background-masked + density-augmented) — are released as separate
-repositories. They are **row-aligned**: each `row_key` exists in all four
-views and refers to the same underlying cell. Each view ships the same
-schema (cell + mask + metadata); only the `cell.bin` content differs.
+| view | repo | cell.bin content |
+|---|---|---|
+| C  | `cp-bg-bench-anon/jump-c`  | raw 5-channel crop |
+| S  | `cp-bg-bench-anon/jump-s`  | background-masked crop |
+| CD | `cp-bg-bench-anon/jump-cd` | density-augmented crop |
+| SD | `cp-bg-bench-anon/jump-sd` | background-masked + density-augmented |
 
-This repository contains view **SD**: background-masked + density-augmented cell crop.
+All four share `row_key`. Join on it to combine views.
 
-## Format
+## Loading
 
-WebDataset tar shards. Each cell occupies one sample with four files sharing a
-key:
+```python
+import json, numpy as np, webdataset as wds
 
-- `{key}.cell.bin` — raw uint8 bytes, shape `(5, 224, 224)`, channel-first.
-  View-specific image content (raw / background-masked / density-augmented).
-- `{key}.mask.bin` — raw uint8 bytes, shape `(2, 224, 224)`, channel-first.
-  Segmentation mask (`[nucleus, cell]`); identical across all four views.
-- `{key}.json` — per-cell metadata (see below).
+URLS = "https://huggingface.co/datasets/cp-bg-bench-anon/jump-sd/resolve/main/data/shard-{{00000..00042}}.tar"
 
-Note: the binary blobs are **raw bytes**, not PNG/NPY. Decode with:
+def decode(s):
+    cell = np.frombuffer(s["cell.bin"], dtype=np.uint8).reshape(5, 224, 224)
+    mask = np.frombuffer(s["mask.bin"], dtype=np.uint8).reshape(2, 224, 224)
+    return cell, mask, json.loads(s["json"])
 
-```python
-import numpy as np
-cell = np.frombuffer(cell_bytes, dtype=np.uint8).reshape((5, 224, 224))
-mask = np.frombuffer(mask_bytes, dtype=np.uint8).reshape((2, 224, 224))
+ds = wds.WebDataset(URLS, shardshuffle=True).shuffle(2000).map(decode)
 ```
 
-## Channels
-
-The 5 channels are stored in this order:
+## Per-cell payload
 
-| index | label | stain |
+| file | shape | dtype |
 |---|---|---|
-| 0 | DNA  | Hoechst 33342 |
-| 1 | AGP  | Phalloidin + WGA |
-| 2 | ER   | Concanavalin A |
-| 3 | Mito | MitoTracker Deep Red |
-| 4 | RNA  | SYTO 14 |
+| `<key>.cell.bin` | (5, 224, 224) | uint8 |
+| `<key>.mask.bin` | (2, 224, 224) | uint8 |
+| `<key>.json`     | -             | json  |
 
-> **Note: this on-disk order is `[DNA, AGP, ER, Mito, RNA]`, which differs
-> from both the JUMP-CP v2 publication order `[DNA, ER, RNA, AGP, Mito]` and
-> the Recursion RxRx order `[DNA, ER, AGP, RNA, Mito, ...]`.** Code that
-> assumes a particular order without checking will feed the wrong stain to
-> downstream channel slots. A machine-readable `channels.json` is provided at
-> the repo root; please use it. The CP-BG-Bench framework resolves channels
-> by name and is robust to per-dataset reorderings.
+`cell.bin` and `mask.bin` are raw bytes. Channel order is in `channels.json`.
 
+### Image channels (cell.bin)
 
-The 2 mask channels are stored in this order:
+| 0 | 1 | 2 | 3 | 4 |
+|---|---|---|---|---|
+| DNA (Hoechst) | AGP (Phalloidin+WGA) | ER (ConA) | Mito (MitoTracker DR) | RNA (SYTO 14) |
 
-| index | label | description |
-|---|---|---|
-| 0 | nucleus | binary nucleus segmentation mask |
-| 1 | cell    | binary whole-cell segmentation mask |
+This order differs from JUMP-CP v2 `[DNA, ER, RNA, AGP, Mito]` and Recursion
+`[DNA, ER, AGP, RNA, Mito, ...]`. Cross-dataset code must remap by name.
 
-Masks are uint8 with values in `{0, 255}` (or another non-zero label).
-A machine-readable `channels.json` is provided at the repo root.
+### Mask channels (mask.bin)
 
+| 0 | 1 |
+|---|---|
+| nucleus | cell |
 
-## Metadata fields
+Binary uint8 in {0, 255}. Identical across all four views.
 
-Each `.json` sidecar contains:
+### JSON metadata
 
-| Field | Type | Description |
+| field | dtype | example |
 |---|---|---|
-| `row_key` | str | Stable cell identifier (`source__date_run__plate__well__tile__id`) |
-| `source` | str | JUMP partner code (e.g. `source_4`) |
-| `batch` | str | Batch identifier (= `source` for JUMP-CP) |
-| `plate` | str | JUMP plate identifier |
-| `well` | str | Well coordinate (e.g. `B02`) |
-| `tile` | str | Site index within well |
-| `id_local` | int | Per-tile cell index |
-| `perturbation` | str | Compound InChIKey (= `Metadata_InChIKey`) |
-| `Metadata_InChIKey` | str | Public PubChem-derivable compound ID |
-| `Metadata_PlateType` | str | `TARGET2` or `COMPOUND` |
-| `nuc_area`, `cyto_area`, `nuc_cyto_ratio` | num | Single-cell QC features |
-| `n_cells_in_fov`, `n_cells_scaled` | num | Local density features |
-
-A flat `metadata.parquet` covering all rows (with `shard_index` and
-`key_in_shard` columns) is provided at the repo root for fast subsetting
-without scanning the tar shards.
-
-## Statistics
-
-- Total cells: 645,390
-- Shards: 43
-- Approximate shard size: ~5.3 GB
-- Plate types: COMPOUND: 445,968, TARGET2: 199,422
-- Sources (JUMP partners): 10
-- Plates: 900
-- Unique perturbations (InChIKeys): 1467
-
-## Splits
-
-A natural held-out batch is `source_4` (21,177 cells), used in the
-CP-BG-Bench cross-batch retrieval evaluation. Filter with:
-
-```python
-ds = ds.filter(lambda r: r["source"] == "source_4")
-```
-
-## Pairing across views
-
-The four CP-BG-Bench JUMP repos share `row_key` exactly. To use a paired
-dataloader, load all four and join on `row_key` (the `metadata.parquet` files
-make this cheap):
-
-- `<anon-org>/jump-c-target2`  (this view: C)
-- `<anon-org>/jump-s-target2`
-- `<anon-org>/jump-cd-target2`
-- `<anon-org>/jump-sd-target2`
-
-## Shard ordering
-
-- Rows are written in the order they appear in the source HF Datasets
-  directory. Shuffle at load time (PyTorch `DataLoader(shuffle=True)` or
-  `webdataset.WebDataset(...).shuffle(N)`); shards are not pre-shuffled.
-- `(shard_index, key_in_shard)` for every row is in `metadata.parquet`.
+| `row_key` | str | `source_4__20210922_Run1__BR00121430__J18__1__27` |
+| `source` | str | `source_4` |
+| `batch`  | str | `source_4` |
+| `plate`  | str | `BR00121430` |
+| `well`   | str | `J18` |
+| `tile`   | str | `1` |
+| `id_local` | int | `27` |
+| `perturbation` | str | `IAZDPXIOMUYVGZ-UHFFFAOYSA-N` |
+| `Metadata_InChIKey` | str | `IAZDPXIOMUYVGZ-UHFFFAOYSA-N` |
+| `Metadata_PlateType` | str | `COMPOUND` or `TARGET2` |
+| `nuc_area`, `cyto_area` | int | nuclear / cytoplasm pixel counts |
+| `nuc_cyto_ratio` | float | derived |
+| `n_cells_in_fov`, `n_cells_scaled` | int, float | local density |
+
+`metadata.parquet` mirrors the JSON fields and adds `shard_index` and
+`key_in_shard` for direct cell lookup.
+
+## Stats
+
+645,390 cells, 43 shards, 10 sources, 900 plates,
+1467 compounds. Plate library: COMPOUND 445,968 cells, TARGET2 199,422 cells.
+Held-out evaluation batch: `source_4` (21,177 cells).
 
 ## License
 
-CC0 1.0 Universal. Underlying JUMP-CP data is released by the JUMP-Cell
-Painting consortium under CC0 1.0.
-
-## Citation
-
-Anonymous submission to NeurIPS 2026 Datasets & Benchmarks Track.
-
-For the underlying JUMP-CP data, please cite the JUMP-CP consortium release.
+CC0 1.0. Anonymous NeurIPS 2026 Datasets and Benchmarks submission; citation
+on acceptance.