Loom52M / README.md
Ruler138's picture
Initial commit
1ba5f71
|
Raw History Blame Contribute Delete
2 kB
metadata
pretty_name: Loom52M
viewer: false
language:
  - en
task_categories:
  - image-to-text
  - text-to-image
tags:
  - biomedicine
  - scientific-figures
  - contrastive-pretraining
  - webdataset
license: other

Loom52M

Training data for BioLoom: Scaling Biomedical Vision-Language Models by Learning from the Structure of Scientific Articles

Loom52M pairs biomedical figures and panels with text derived from the scientific literature. Whole figures retain their original captions. Panel text combines a visual description with a contextual caption grounded in the parent figure caption and relevant article paragraphs. The hierarchy preserves article, figure, and panel relationships for contrastive learning.

📦 Dataset layout

Path Contents
train/ 4,525 hierarchical WebDataset tar shards, 52,437,129 image–text pairs
manifests/ Training figure index
dataset_stats.json Training corpus statistics

Released models and paper evaluation data are hosted separately.

📥 Download and use

Download the metadata and one shard first:

hf download ChartFoundation/Loom52M --repo-type dataset   --include 'dataset_stats.json' 'train/hier-w00-000000.tar'   --local-dir Loom52M
tar -tf Loom52M/train/hier-w00-000000.tar | head

Download the full training split with:

hf download ChartFoundation/Loom52M --repo-type dataset \
  --include 'train/*.tar' 'dataset_stats.json' --local-dir Loom52M

Each WebDataset sample groups an image, caption, and JSON metadata under the same key. Keep the tar shards intact and pass the train/ shard list to the hierarchical loader in the training implementation. The shards total approximately 2.42 TB. Use BioLoom-EvalSuite to evaluate trained models.