BioLoom-EvalSuite / README.md
Ruler138's picture
Initial commit
a220039
|
Raw History Blame Contribute Delete
3.1 kB
metadata
pretty_name: BioLoom Evaluation Suite
viewer: false
language:
  - en
task_categories:
  - image-classification
  - image-to-text
  - text-to-image
tags:
  - biomedical
  - benchmark
  - zero-shot
  - retrieval
license: other

BioLoom Evaluation Suite

Evaluation data and protocols for BioLoom: Scaling Biomedical Vision-Language Models by Learning from the Structure of Scientific Articles.

🧪 Benchmarks and evaluation

Benchmark Scope Contents
Zero-shot classification 67 tasks, 43 sources, 10 domains Fixed prompts, a task manifest, and evaluation data
BioLoomBench 15,000 figures, 15 domains Retrieval with concise and original captions
BioMedFlickr 7,185 image–caption pairs Test gallery and both relevance protocols

BioLoomBench

BioLoomBench is a biomedical image–text retrieval benchmark with 15,000 whole figures from 6,990 articles published or posted in 2026 and collected through bioRxiv and Europe PMC Open Access. It covers 15 biomedical domains with 1,000 figures each, spanning biological images, mixed image–chart figures, diagrams, and chart-only figures. Each figure has a concise visual caption and the source caption from the article.

How to use

To reproduce the main paper results, download the 30 WebDataset shards in main-data/ and the metadata file.

hf download ChartFoundation/BioLoom-EvalSuite --repo-type dataset \
  --include 'bioloom-retrieval/main-data/*.tar' 'bioloom-retrieval/metadata.jsonl' \
  --local-dir BioLoom-EvalSuite
  1. Read the JPEG and JSON record for each figure from the tar shards. Use metadata.jsonl for the released figure order, domain labels, and caption fields.
  2. Evaluate concise_caption and source_caption separately using the model’s native context length. For the additional source-caption setting, use the smaller of 77 tokens and the model’s native limit.
  3. Rank all 15,000 candidates in each direction. Use the figure’s paired caption or image as the relevant item, and report Recall@1, Recall@10, and Recall@100 for image-to-text and text-to-image. For a domain or visual-content breakdown, restrict the queries and keep the full gallery.

The main paper reports BioLoomBench retrieval results using main-data/ and metadata.jsonl. Appendix I (Fine-grained and cross-level retrieval) uses panel-retrieval/ for panel-level and cross-level tasks, including panel-description-to-figure retrieval. See the panel retrieval guide.

Model checkpoints · Training corpus · Implementation