---
pretty_name: "COVER-Fish Reference-Update Benchmark"
license: other
task_categories:
  - image-classification
tags:
  - retrieval
  - reference-gallery
  - benchmark
  - data-centric-ai
  - biodiversity-informatics
  - reproducibility
size_categories:
  - 100K<n<1M
---

# COVER-Fish Reference-Update Benchmark

This benchmark studies how reference-corpus updates change a frozen recognition
system, instantiated on fine-grained fish identification. It publishes the
versioned control plane behind COVER-Fish: row-level manifests, gallery states,
taxonomy, tensor bindings, transition evidence, protocols and dependency locks.

The benchmark does **not** duplicate the 83 GB Full Payload Archive. Large source
archives and frozen tensors remain in the immutable release at DOI
[`10.57967/hf/9776`](https://doi.org/10.57967/hf/9776), revision
`4e437b6a2bf5f9a12a200bbe3a93411fe713db1f`. The lock in
`dependencies/DEPENDENCIES.tsv` binds every required object by path, byte count
and complete SHA-256.

## Canonical tracks

1. **Gallery-state replay** — hold encoder, query roster and aggregation rule
   fixed while comparing R0, R1, R2, R3, R4, +USFWS, +ANGFA and +Commons.
2. **Aggregation replay** — hold embeddings and gallery membership fixed while
   comparing species-centroid, instance-max, source-separated or another fully
   declared aggregation rule.

Both tracks report query-micro and species-macro top-1/top-5/top-20 together
with wrong-to-right and right-to-wrong transitions. This release does not create
an acquisition leaderboard over the public QT26-QC labels.

## What to download

| Goal | This repository | Fixed dependency |
|---|---|---|
| Inspect states, licences and protocols | manifests, views, ledgers | none |
| Run the 16-query Level A fixture | control files and example | CORE (~0.50 GB) |
| Replay rankings or aggregation | state definitions and protocol | CORE |
| Re-encode source images | manifests and licence lists | selected S0–S4 archives |
| Evaluate a new recognition model | QT26-QC dataset | none from this repository |

The evaluation roster is independently published as
[`COVER-Fish/QT26-QC`](https://huggingface.co/datasets/COVER-Fish/QT26-QC).
Its E0 copy remains a dependency here so the complete historical replay is
closed under the same fixed snapshot.

## Minimal Level A replay

```bash
python tools/download_dependency.py CORE --output dependencies/downloads
tar --zstd -xf dependencies/downloads/coverfish-rev023-core-metadata-index-code-rc2-20260714.tar.zst \
  -C dependencies/core
python examples/minimal_reproduction.py dependencies/core/index
```

The fixture uses NumPy, frozen query embeddings and frozen species prototypes;
it requires no pixels, network access after download, PyTorch or encoder run.

## Licence-aware selection

`rows-by-licence/` contains row-ID lists for CC0/public-domain, CC BY, CC BY-SA,
non-commercial, permission-required and pointer-only records. These lists make
the row-level rights ledger operational: after obtaining and extracting a source
archive, `tools/extract_rows.py` can select the requested rows.

The lists are **not** separately downloadable image packs. In particular,
Commons is one 66.6 GB compressed stream transported in nine parts; selecting a
Commons subset requires obtaining and extracting that stream first. Row-level
terms remain authoritative.

## Dependency verification

Anyone can run `python tools/check_dependencies.py` to compare the dependency
lock with the fixed Hub revision. The tool uses LFS content oids when available,
fully hashes small non-LFS files, and can fall back to two fixed byte-range
hashes without calling them a complete-file SHA-256. It never executes strings
from TSV files. No fixed monitoring frequency or continuing verification service
is promised.

## Citation

- Use this benchmark's DOI when using gallery states, embeddings, ledgers or
  replay protocols.
- Use the QT26-QC DOI when using the query set or reporting benchmark results.
- Use the COVER-Fish paper when using its method or scientific conclusions.
- Cite `10.57967/hf/9776` and name the component when directly using S0–S4,
  D0, E0 or CORE bytes.

If several objects are used, cite the objects actually downloaded or used to
produce the result.