---
pretty_name: "QT26-QC: A Temporally Bounded Fish Recognition Evaluation Roster"
license: other
task_categories:
  - image-classification
tags:
  - fish-recognition
  - fine-grained-recognition
  - biodiversity-informatics
  - benchmark
  - temporal-evaluation
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet
---

# QT26-QC

QT26-QC is a fixed evaluation roster for fine-grained fish recognition. Its
canonical `test` split contains **6,719 natural-photo queries across 3,121
FishBase 25.04 taxa**, with encoded image bytes, labels, provenance, row-level
licence and attribution, and image SHA-256 values.

```python
from datasets import load_dataset

test = load_dataset("COVER-Fish/QT26-QC", split="test")
```

## Why this benchmark is complementary

QT26-QC complements existing fish benchmarks along the temporal axis: FishNet
provides a larger natural-photo roster, Fish-Vista provides museum specimens,
and QT26-QC provides natural photographs whose recorded observation times all
fall after a declared cutoff. The canonical window is **1 March 2026 02:31 UTC
through 23 June 2026**.

This boundary must be interpreted relative to each evaluated model's own
training cutoff. It does not prove that source pixels never appeared earlier on
the web or guarantee disjointness from any model's pretraining data.

## Canonical scoring set

The sole scoring split is `test` (6,719 rows). A 6,734-row construction parent
roster is retained under `metadata/` for lineage only; the 15 Q-INT
observation-overlap rows are listed under `construction/` and are not present in
the image split. The immutable E0 archive in the Full Payload Archive also
contains the same canonical 6,719 rows.

The required submission format and fixed-denominator metrics are specified in
[`evaluation/protocol.md`](evaluation/protocol.md). The included scorer reports
query-micro and species-macro top-1/top-5/top-20 with a 20,000-replicate
target-species-cluster bootstrap.

## Two byte-identical access paths

- This repository stores the canonical images in Parquet for one-line loading
  and Dataset Viewer access.
- `archive-access/download_original_e0.py` retrieves the immutable E0
  `.tar.zst` from the Full Payload Archive at DOI
  [`10.57967/hf/9776`](https://doi.org/10.57967/hf/9776), revision
  `4e437b6a2bf5f9a12a200bbe3a93411fe713db1f`.

`metadata/row-correspondence.tsv` binds every Parquet row to its E0 archive
member. `schema/verify_bytes.py` independently hashes the encoded bytes stored
in Parquet; the build receipt records the full 6,719-row cross-check.
`archive-access/check_dependencies.py` can be run manually to verify that the
fixed E0 dependency still resolves to the recorded LFS content SHA-256. This is
a user-invoked capability, not a continuing monitoring service. If an API omits
the LFS oid, the tool can compare two fixed byte-range hashes while clearly
distinguishing that check from a complete-file SHA-256.

## Rights

This dataset uses `license: other` because the images retain their source
licences. Every row records `license_normalized`, source URLs and attribution.
The split contains CC BY-NC, CC BY, CC0, CC BY-NC-ND, CC BY-NC-SA and CC BY-SA
material. The metadata and evaluation code do not relicense the underlying
images; consult each row before reuse beyond evaluation.

## Citation

- Cite the QT26-QC dataset DOI shown on this repository when using its queries
  or reporting benchmark results.
- Cite the COVER-Fish Reference-Update Benchmark when using gallery states,
  frozen embeddings or transition ledgers.
- Cite the COVER-Fish paper when using its method or scientific conclusions.
- Cite DOI `10.57967/hf/9776` when directly using the E0 archival object.

See [`CITATION.cff`](CITATION.cff) for creator metadata.