README.md

---
license: cc-by-4.0
pretty_name: "PHOEBI"
language:
- en
task_categories:
- image-classification
task_ids:
- multi-label-image-classification
size_categories:
- 100K<n<1M
tags:
- bacteria
- microscopy
- phase-contrast
- multi-label-classification
- compositional-generalization
- open-set-recognition
- novel-class-discovery
- fine-grained-recognition
- benchmark
configs:
- config_name: phoebi6
  default: true
  data_files:
  - split: train
    path: data/phoebi6/train-*.parquet
  - split: val
    path: data/phoebi6/val-*.parquet
  - split: test
    path: data/phoebi6/test-*.parquet
- config_name: phoebi4
  data_files:
  - split: train
    path: data/phoebi4/train-*.parquet
  - split: val
    path: data/phoebi4/val.parquet
  - split: test
    path: data/phoebi4/test.parquet
---

# PHOEBI

**P**hase-contrast **O**ptical b**E**nchmark for **B**acterial **I**dentification — a benchmark and framework for open-world identification of mixed bacterial cultures from optical phase-contrast microscopy.

## At a glance

The release contains two subsets that share the same imaging pipeline:

| Subset | Species | Combinations | Images |
|--------|---------|--------------|--------|
| **PHOEBI-6** (primary) | 6 (`bs`, `bt`, `fj`, `ka`, `mx`, `pf`) | 40 | ~120,000 |
| **PHOEBI-4** (legacy)  | 4 (`b`, `f`, `k`, `p`)                 | 14 | ~14,000  |

Both subsets:

- 1024×1024×3 JPEG images at 1000× total magnification (100× oil-immersion, NA 1.25).
- Random 80/10/10 split (PHOEBI-6 also ships a leave-combinations-out (LCO) split) at seed 1337.
- Croissant 1.0 metadata in `croissant.json`.
- Released under CC BY 4.0.

The 4-class subset was collected under a separate microscopy session and is included to support cross-session replication of the compositional-collapse finding (paper §4.5, Appendix E.4).

## Species (PHOEBI-6)

| Code | Latin name | Gram | Motility | Cell length |
|------|------------|------|----------|-------------|
| `bs` | *Bacillus subtilis*       | + | peritrichous flagella | 4–10 µm |
| `bt` | *Bacillus thermoamylovorans* | + | peritrichous flagella | ~4 µm |
| `fj` | *Flavobacterium johnsoniae* | − | gliding | 5–10 µm |
| `ka` | *Klebsiella aerogenes*    | − | peritrichous flagella | 1–3 µm (encapsulated) |
| `mx` | *Myxococcus xanthus*      | − | gliding | 5–10 µm |
| `pf` | *Pseudomonas fluorescens* | − | polar flagella | 1.5–3 µm |

## Directory layout

```
.
├── croissant.json                 # Croissant 1.0 metadata (core + RAI)
├── LICENSE                        # CC BY 4.0
├── README.md                      # this file
├── splits.json                    # PHOEBI-6 random + LCO split definitions (seed 1337)
├── images.tar.gz                  # PHOEBI-6: ~120,000 1024×1024 images
├── splits_4class.json             # PHOEBI-4 random split definition (seed 1337)
└── images_4class.tar.gz           # PHOEBI-4: ~14,000 1024×1024 images
```

After extraction the layout is:

```
images/                            # PHOEBI-6
├── bs/                            # singletons
├── bs_ka/                         # pairs (combo encoded by underscore-joined codes)
├── bs_ka_fj/                      # triples
└── bs_bt_mx_ka_fj_pf/             # six-species mixture

images_4class/                     # PHOEBI-4
├── b/                             # 4-class singletons
├── b_f/                           # 4-class pairs
├── b_f_k/                         # 4-class triples
└── b_f_k_p/                       # 4-class quadruple
```

Each image is a 1024×1024×3 RGB JPEG. The PHOEBI-6 label vector is the multi-hot indicator over `[bs, bt, fj, ka, mx, pf]`; PHOEBI-4 over `[b, f, k, p]`. Example: `bs_ka_fj` → `[1, 0, 1, 1, 0, 0]`.

## Splits

`splits.json` (PHOEBI-6) and `splits_4class.json` (PHOEBI-4) define the protocols.

**Random 80/10/10.** Image-level split with a fixed seed, intended for in-distribution closed-set characterisation.

**Leave-combinations-out (LCO), seed 1337** *(PHOEBI-6 only)*. Holds out entire species combinations under three constraints: combination disjointness (held-out combinations never appear in train or val), species coverage (every species appears in at least one trained-on combination, so the protocol tests compositional generalization rather than novel-class detection), and order coverage (held-out set spans a range of combination orders).

The PHOEBI-6 LCO held-out combinations at seed 1337 are: `bt`, `bs_pf`, `ka_fj`, `bs_mx_fj`, `bs_ka_pf`, `mx_ka_fj`, `bs_bt_ka_fj`, `bs_mx_fj_pf`, `bs_bt_mx_ka_fj_pf`.

## Wet-lab protocol

Cultures were inoculated from glycerol stocks into nutrient broth (8 g L⁻¹), sterilised by autoclave (121 °C, 15 min), and grown at 30 °C with orbital shaking at 250 rpm for 72–120 hours. Once each culture reached its characteristic growth stage, samples were mounted on glass slides and imaged on an inverted phase-contrast microscope (100× oil-immersion, NA 1.25). For PHOEBI-6 three pairwise combinations (`bs`+`bt`, `bt`+`fj`, `ka`+`pf`) and the five-species combinations were not collected and are absent from the release.

## Tasks supported

1. **Multi-label species presence detection** in mixed cultures.
2. **Compositional generalization** (LCO protocol).
3. **Open-set rejection** (leave-one-class-out evaluation; see paper §4.4).
4. **Novel-class discovery** (multi-label compositional NCD; see paper §4.4).

## Tasks NOT supported

- **Proportion / composition estimation.** Ground-truth species ratios are unavailable; growth rates differ across species and the wet-lab procedure records only species presence, not relative abundance.
- **Clinical pathogen identification.** No stained-smear protocol; the dataset must not be used for clinical decisions without independent validation.

## Loading the data

A minimal Python loader using only the standard library:

```python
import json
from PIL import Image
import numpy as np

# Assume images.tar.gz has been extracted next to splits.json
with open("splits.json") as f:
    splits = json.load(f)

class_names = splits["class_names"]   # ['bs', 'bt', 'fj', 'ka', 'mx', 'pf']

for entry in splits["splits"]["train"]:
    img = np.array(Image.open(entry["path"]).convert("RGB"))   # 1024 x 1024 x 3
    label = np.asarray(entry["label"], dtype=np.int64)         # 6-vector multi-hot
    combo = entry["combo"]                                     # combination token, e.g. "bs_ka"
    # ... your pipeline ...
```

For the LCO protocol, use the helper at `baselines.supervised_multilabel_heldout.select_heldout(seed=1337)` in the companion code release for an apples-to-apples held-out set.

## Citation

```bibtex
@misc{phoebi2026,
  title        = {{PHOEBI}: An Open-World Benchmark for Bacterial Identification in
                  Phase-Contrast Microscopy},
  author       = {Anonymous Authors},
  year         = {2026},
  note         = {Under review at the NeurIPS 2026 Datasets and Benchmarks Track},
  howpublished = {OpenReview: \url{https://openreview.net/PLACEHOLDER}},
}
```

(Author and institution information will be added on acceptance.)

## Contact

During the double-blind review period, please use the OpenReview discussion thread for the corresponding submission. Public contact information will be added on acceptance.