GitHub - jonathanlyardley/WildlifeBench: WildlifeBench is an open benchmark and study that tests AI-assisted camera-trap labelling routes on GBIF-mediated media, in order to assist decision-support workflow, showing which routes are useful, and for what species or continent before human-supervised species labelling.

Practical summary

Hugging Face image/dataset repository: https://huggingface.co/datasets/JonathanYardley/wildlifebench-gbif-ebbe-2026-data
The cropped/uncropped 246 camera-trap bursts / 82 species JPEG media live here, not in this GitHub repo.
Media package: 2,010 cropped/uncropped JPEG derivatives with per-media attribution, licence URLs, EXIF-stripping notes and checksums, hosted on the Hugging Face dataset above.
Audit trail for the run: exact prompt (PROMPTS.md), per-run manifests with Langfuse trace IDs and the exact run code (results/run_outputs/, model_pipeline/), and reproduction steps (REPRODUCE.md).
Per-event model outputs are restricted to the 246 events;
No DOI or GBIF derived-dataset registration is claimed here.

Read SUBMISSION.md and docs/METHODOLOGY.md.
Set WBGBIF_DATA_PACKAGE to the local data-package folder, or place that folder beside this repo as data_package.
Validate the release with python scripts/validate_release.py.
Rebuild the displayed table with python scripts/build_submission_table.py.
Inspect results/speciesnet_overlap/application_safe_results.md.
For deeper audit, read PROMPTS.md, the per-run manifests and raw outputs in results/run_outputs/, and the exact run code in model_pipeline/. Full reproduction steps are in REPRODUCE.md.
Open or download gbif_award_preview.html (repo root) for the self-contained visual preview of the same displayed comparison.
Review data rights in DATA_AND_MEDIA_RIGHTS.md and the data package files.

The no-cost route uses cached outputs and does not require paid API credentials.