Gumpest/AutoV-data

AutoV Attention-Map Training Data

This dataset repository contains the training metadata and derived attention-map images used by AutoV. It does not contain the original source images.

Contents

Training data

• metadata/llava_v1_5_mix100k_reward.json: 100,000 reward-training records.
• metadata/llava_v1_5_mix100k_filter.json: the corresponding filtered records.
• metadata/llava_v1_5_mix1k_*.json: the 1,000-record development subset.
• archives/attnmap100K/: 20 tar archives containing 400,000 attention maps.
• archives/attnmap1K/: 4 tar archives containing 4,000 attention maps.
• SHA256SUMS: integrity checks for metadata and archives.

Inference data

• inference-data/mmvet/: 800 MMVet attention-map images (200 image names across CLIP ViT-L/14@336 layers 15, 20, 22, and 23). The directory layout matches LLaVA/playground/data/api_eval/mmvet.

The 100K metadata references 25,000 COCO, 20,000 Visual Genome, 20,000 GQA, 20,000 OCR-VQA, and 15,000 TextVQA source-image paths. Each record has four derived attention maps from CLIP ViT-L/14@336 layers 15, 20, 22, and 23.

Restore the original AutoV data layout

Run from the downloaded repository root:

mkdir -p data/attnmap100K data/attnmap1K
for archive in archives/attnmap100K/*.tar; do
  tar -xf "$archive" -C data/attnmap100K
done
for archive in archives/attnmap1K/*.tar; do
  tar -xf "$archive" -C data/attnmap1K
done
cp metadata/* data/
sha256sum -c SHA256SUMS

Licensing note

The metadata refers to COCO, Visual Genome, GQA, OCR-VQA, TextVQA, and LLaVA data. Their original licenses and terms continue to apply. Redistribution of the derived attention-map images was reviewed against the applicable source terms before publication.