Phase-contrast Optical bEnchmark for Bacterial Identification: 120,000 phase-contrast microscopy images of 40 combinations of six rod-shaped bacterial species, for recognising which species a culture contains, including combinations and species never seen in training.

Paper: PHOEBI: An Open-World Benchmark for Multi-Label Bacterial Identification in Phase-Contrast Microscopy, NeurIPS 2026, Track on Evaluations and Datasets.

Code: github.com/eternal-f1ame/phoebi

Project page: phoebi-benchmark.vercel.app

Config	Species	Combinations	Images	Image size
`phoebi6` (default)	6	40	120,000	1024 × 1024
`phoebi4` (legacy)	4	14	14,000	square, about 1,000–1,800 px

phoebi4 is an earlier collection, cultured in a separate batch and imaged in a separate session on the same instrument; the paper uses it to replicate its finding that fine-tuned classifiers collapse on unseen combinations. Its codes b, f, k and p are Bacillus subtilis, Flavobacterium johnsoniae, Klebsiella aerogenes and Pseudomonas fluorescens, the species coded bs, fj, ka and pf in phoebi6.

Field	Description
`image`	RGB phase-contrast image (JPEG) at 1000× total magnification
`labels`	the species present, as class labels (names below)
`combination`	the culture combination, species codes joined by `_`, e.g. `bs_ka_fj`
`lco_split`	`phoebi6` only: `train`, `validation` or `heldout` under the leave-combinations-out protocol

Species

Code	Species	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

phoebi6 covers all 6 singletons, 12 of the 15 pairs, 15 of the 20 triples, 6 of the 15 quadruples and the full six-species mixture, 3,000 images each.

Splits and protocols

Random 80/10/10 (train / validation / test). Each combination's images are assigned in acquisition order, the first 80% to training, the next 10% to validation and the last 10% to test, and the three crops of a frame always share a split. Use it for in-distribution characterisation.
Leave-combinations-out (LCO), seed 1337 (phoebi6, column lco_split). Nine whole combinations are held out: bt, bs_pf, ka_fj, bs_mx_fj, bs_ka_pf, mx_ka_fj, bs_bt_ka_fj, bs_mx_fj_pf and bs_bt_mx_ka_fj_pf (27,000 images). The other 31 are split 90/10 into training and validation (83,700 / 9,300). Every species still appears in training, so LCO measures compositional generalisation: recognising known species in mixtures never seen during training.
Development folds (lco_dev_folds.json). Five folds over the trained-on combinations, for model selection; the nine held-out combinations are then used once, for reporting.

Every image of a combination carries the same labels, so held-out metrics compare combinations rather than images, and the paper reports bootstrap intervals over combinations.

Loading

from datasets import load_dataset

ds = load_dataset("sochastic/PHOEBI", "phoebi6")                  # random split
species = ds["train"].features["labels"].feature.names            # ['bs', 'bt', 'fj', 'ka', 'mx', 'pf']

# Leave-combinations-out: take all 120,000 images and split them by lco_split.
full = load_dataset("sochastic/PHOEBI", "phoebi6", split="train+validation+test")
lco = {}
for name in ("train", "validation", "heldout"):
    lco[name] = full.filter(lambda s, name=name: s == name, input_columns="lco_split")

splits.json, splits_lco.json and splits_4class.json list every image by the path stored in image["path"], with its multi-hot label vector; the code reads these, and its tools/fetch_dataset.py writes every image to that path. croissant.json holds the Croissant metadata.

Collection

Each species was grown separately from a glycerol stock in nutrient broth (30 °C, 250 rpm, 72–120 h), inspected under the microscope and confirmed pure. A combination was made only at acquisition time, by mixing the verified suspensions at a controlled volume ratio and mounting the mixture immediately as an unstained wet mount, so the species were never co-cultured and every label is verified at culture level. Videos were recorded under phase contrast at 1000× total magnification (100× oil immersion, NA 1.25) with a colour camera, 1,000 frames were sampled from each combination's footage, and each frame gave three random square crops resampled to 1024 × 1024. A species can still miss an individual field by sampling; the paper bounds this rate at under 5% on average.

Intended use and limitations

PHOEBI is a research benchmark for multi-label recognition, compositional generalisation, open-set rejection and novel-class discovery. It records which species are present, not their abundance, and it was acquired on a single instrument. It is not intended for clinical diagnosis.

Citation

@inproceedings{baranwal2026phoebi,
  title         = {{PHOEBI}: An Open-World Benchmark for Multi-Label Bacterial Identification in Phase-Contrast Microscopy},
  author        = {Baranwal, Aaditya and Hasan, Md Jahid and Vyas, Shruti},
  booktitle     = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
  year          = {2026},
  eprint        = {2606.22890},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}