README.md

---
license: other
task_categories:
- image-classification
size_categories:
- 1K<n<10K
---

# multi-species benchmark

Photographs where 2+ species appear in the same frame. Designed to evaluate
multi-label species identification and steering capabilities of biological
vision-language models. Two sources, unified into one parquet schema.

## Sources

### `inat21_multilabel` (299 rows, 147 images)

In-distribution: drawn from [iNat21](https://github.com/visipedia/inat_comp/tree/master/2021)
validation images that already carry an iNat-supplied primary label. We use
[InternVL3-AWQ](https://huggingface.co/OpenGVLab/InternVL3-38B-AWQ) to surface
images that also contain a *secondary* species, then
[BioCLIP2](https://huggingface.co/imageomics/bioclip-2) to propose
zero-shot candidates, then two human reviewers (ariana, dan) to accept/reject
and confirm the secondary species. Each accepted image contributes a row for
the primary organism (with iNat ground-truth taxonomy) and a row per
secondary organism (with reviewer-confirmed taxonomy).

### `lila_camera_trap` (884 rows, 442 images)

Out-of-distribution: trail-camera images from
[LILA BC](https://lila.science/lila-bc/) where the source labels (image-level,
multiple species in the same frame) have been validated by the LILA
community. Filtered to image-level multi-species cases with fully species-
resolved taxa, humans + domestic animals excluded.

## Schema

| field | type | notes |
|---|---|---|
| `image_id` | str | unique key |
| `source_benchmark` | str | `inat21_multilabel` or `lila_camera_trap` |
| `source_dataset` | str | finer-grained source label (e.g. `iNat-2021-val`, `Nkhotakota Camera Traps`) |
| `source_url` | str | original URL when available (LILA); null for iNat |
| `image` | struct (bytes, path) | JPEG bytes inline |
| `n_species_in_image` | int32 | count of distinct species in this image |
| `species_idx` | int32 | 0 for primary, 1..N for secondaries |
| `role` | str | `primary` / `secondary` for iNat21; null for camera trap |
| `bbox_px` | list&lt;int32&gt; | pixel coords [x1,y1,x2,y2] of the organism; null for camera trap |
| `kingdom` ... `species` | str | 7 taxonomic ranks (iNat convention: capitalized ranks, lowercase species epithet) |
| `common_name` | str | when available |
| `taxonomy_source` | str | `inat21_ground_truth` / `bioclip_top_k_index_N` / `free_text` / `lila_camera_trap` |
| `reviewer` | str | reviewer name for iNat rows; null otherwise |

## Statistics

### `inat21` — 299 rows, 147 images

Species-per-image distribution: 2: 144, 3: 2, 5: 1

Top source datasets:
- `iNat-2021-val`: 299 rows

Top species (by row count):
- `purpuratus`: 8 rows
- `canadensis`: 3 rows
- `pluchei`: 3 rows
- `muricata`: 3 rows
- `vatia`: 3 rows
- `herbacea`: 2 rows
- `iguana`: 2 rows
- `ravilla`: 2 rows

### `camera_trap` — 884 rows, 442 images

Species-per-image distribution: 2: 442

Top source datasets:
- `Nkhotakota Camera Traps`: 668 rows
- `ENA24`: 158 rows
- `Orinoquia Camera Traps`: 44 rows
- `Caltech Camera Traps`: 14 rows

Top species (by row count):
- `sylvaticus`: 235 rows
- `cynocephalus`: 233 rows
- `floridanus`: 76 rows
- `brachyrhynchos`: 68 rows
- `strepsiceros`: 41 rows
- `grimmia`: 34 rows
- `pygerythrus`: 30 rows
- `africanus`: 20 rows


## Citation

If you use this benchmark, please cite the underlying sources:

- iNat21: <https://github.com/visipedia/inat_comp/tree/master/2021>
- LILA BC: <https://lila.science/lila-bc/>
- InternVL3: <https://huggingface.co/OpenGVLab/InternVL3-38B-AWQ>
- BioCLIP2: <https://huggingface.co/imageomics/bioclip-2>

## License

iNat21 photos: per-observation licenses (typically CC BY-NC). LILA BC: varies
by dataset; see <https://lila.science/datasets/>. Use of this benchmark for
*research / evaluation* is intended; review the upstream licenses before any
redistribution.