# Clustering the BioTrove Dataset — Data

来源：https://www.kaggle.com/competitions/biotrove-clustering/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

## Files
* **metadata.csv** — metadata for the released subset
  * `hash_id` — anonymized identifier; matches image filenames in `images.zip`
  * `family` — the only taxonomic label provided. Genus and species are hidden for scoring.

* **images.zip** — ~50,000 biodiversity images. Each file is named `<hash_id>.jpg` corresponding to rows in `metadata.csv`.

* **sample_submission.csv** — example submission file in the correct format
  * `hash_id` — from `metadata.csv`
  * `species_cluster` — your cluster assignment at the species level (e.g., 1-S, where S is number of species)
  * `genus_cluster` — your cluster assignment at the genus level  (e.g., 1-G, where G is number of genus)
  * `family_cluster` — your cluster assignment at the family level  (e.g., 1-F, where F is number of families). Since families are given, you can also use the actual names here.

## Predicting cluster labels
You are not predicting explicit taxonomic labels. Instead, you assign each image to clusters at three levels (species, genus, family). Your clustering will be evaluated against hidden ground truth labels using **Normalized Mutual Information (NMI)**. 

## Data distribution 
This data was prepared from BioTrove using random sampling, with the following criteria:

* Each species has a minimum number of 10 images (median=14)
* Each genus a minimum of 2 species (median=3)
* Each family includes minimum of 3 genera (median=5)

The result is not perfectly balanced, but ensures meaningful coverage across taxonomic ranks for clustering.

## License and Attribution
The BioTrove dataset is publicly available under Creative Commons licenses via iNaturalist.  
This competition uses a ~50,000-image subset derived from BioTrove, with hashed IDs replacing original filenames.  

### BioTrove
Yang et al. (2024). BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity.
NeurIPS 2024 Datasets and Benchmarks Track.
[Dataset page](https://huggingface.co/datasets/BGLab/BioTrove) · [Paper](https://openreview.net/forum?id=DFDCtGQs7S)