---
license: cc-by-4.0
language:
- en
pretty_name: cp-bg-bench preview (jump)
size_categories:
- 1K<n<10K
task_categories:
- image-classification
tags:
- cell-painting
- microscopy
- biology
- benchmark
configs:
- config_name: crops
  data_files:
  - split: train
    path: crops/data-*.arrow
- config_name: crops_density
  data_files:
  - split: train
    path: crops_density/data-*.arrow
- config_name: seg
  data_files:
  - split: train
    path: seg/data-*.arrow
- config_name: seg_density
  data_files:
  - split: train
    path: seg_density/data-*.arrow
---

# cp-bg-bench preview — jump

Compact preview of the **jump** dataset from the cp-bg-bench
benchmark. Stratified subset of the full release; designed so that the
held-out-batch perturbation-recall and cp_measure-prediction evals can
be reproduced end-to-end against this small slice alone.

| | |
|---|---|
| Cells | 913 |
| Wells | 69 |
| Perturbations | 33 |
| Held-out batch | `source_4` |
| Views | crops, crops_density, seg, seg_density |

## What's in this repo

```
crops/                # HF dataset, 224×224×C uint8 cell crops + masks
crops_density/        # crops + 4 corner density patches
seg/                  # mask-applied cell crops
seg_density/          # mask + density patches
metadata/             # perturbations parquet + quality filter report
build_info.json       # full provenance (seed, source pipeline commit, selection stats)
```

## Schema (per cell row)

```
row_key, source, plate, well, tile, id_local                              # identity
nuc_area, cyto_area, nuc_cyto_ratio, n_cells_in_fov, n_cells_scaled        # per-cell QC
mask:  large_binary  # (2, 224, 224) uint8 -> NucMask, CellMask
cell:  large_binary  # (C, 224, 224) uint8 fluorescence channels
Metadata_JCP2022, Metadata_InChIKey, Metadata_PlateType                    # source metadata
perturbation, batch, treatment, Metadata_Perturbation                      # benchmark axes
```

## Selection recipe

1. All control wells, capped to a small budget.
2. Treated wells whose perturbation appears in **both** the held-out
   batch and at least one non-held-out batch (so cross-batch retrieval
   is reproducible). Per chosen perturbation: one held-out well + one
   non-held-out well.
3. Per-well cell cap to keep total budget near 1,000 cells.

Seed = 42. See `build_info.json` for exact stats.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("cp-bg-bench-anon/jump-sample", name="crops", split="train")
print(ds)
print(ds[0]["row_key"], ds[0]["perturbation"])
```

## Full dataset

The corresponding full-size release is hosted separately.

## License

CC0 1.0. Anonymous NeurIPS 2026 Datasets and Benchmarks submission; citation on acceptance.