MeasL-150K-V1

MeasL-150K-V1 is the official PRSIMVL instruction-tuning corpus for measurement-grounded visual reasoning. It is aligned with the MeasL-Bench-V1 evaluation protocol and released for reproducible non-commercial research.

• Hugging Face: kepeng/MeasL-150K-V1
• Project page: https://kepengxu.github.io/projects/prism-vl/
• Paper: https://arxiv.org/abs/2605.11727

At A Glance

Release Contents

Path Convention

Training records use local relative image paths under:

training_data/image/

The Hugging Face release stores images as uncompressed tar shards to avoid the Hub limit of 10,000 files per directory. Each tar shard preserves paths such as image/measl_150k_v1_000000.png; after extraction, the training configs can resolve image paths directly.

Recommended release-layout download and extraction:

# Run from the parent directory that contains release_neurips_version_code/.
hf download kepeng/MeasL-150K-V1 \
  --repo-type dataset \
  --local-dir release_neurips_version_code/training_data

cd release_neurips_version_code/training_data
for shard in image_shards/*.tar; do
  tar -xf "$shard"
done

After extraction, images are materialized at release_neurips_version_code/training_data/image/, so JSONL paths such as training_data/image/measl_150k_v1_000000.png resolve from the release root.

If you use the Hugging Face dataset repository as a standalone folder, extract the shards and add a lightweight path bridge:

cd MeasL-150K-V1
for shard in image_shards/*.tar; do
  tar -xf "$shard"
done
mkdir -p training_data
ln -s ../image training_data/image

The shards are tar-only archives, not gzip or xz archives. This keeps packaging deterministic and avoids CPU-heavy compression during upload or extraction.

Packaging Image Shards

If you need to recreate the release shards from a local image/ directory:

cd release_neurips_version_code/training_data
python create_image_tar_shards.py --shard-size 5000 --overwrite

This creates 10 uncompressed tar files under image_shards/ for the 48,000 release images. Upload train-measl-150k-v1.jsonl, README.md, create_image_tar_shards.py, and image_shards/; do not upload the raw image/ directory.

Dataset Construction Summary

The released corpus was built from approximately 700K auto-annotated candidates. After quality scoring, 518,433 records remained. The final 152,517 examples were balanced by source and question structure, packaged with release-local images, and checked to remove strict benchmark overlaps.

JSONL Format

Each row follows the multimodal conversation format used by the release training configs. Typical fields include:

Dataset Registration

The release registry maps these aliases to the release-local training JSONL:

• MEASL/TRAIN_150K_V1
• MEASL/TRAIN_150K_V1_EXTRA

Training launch scripts are documented in the main release README and live under:

configs/qwen3_vl_150k_llmmeta_vit_proxy/

Quality And Integrity

• Benchmark decontamination is applied before release.
• The corpus is intended for reproducible instruction tuning, not benchmark evaluation.
• Modified or expanded variants should be published under a new semantic version, for example MeasL-150K-V2.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

You may use, share, and adapt the dataset for non-commercial research and educational purposes, provided that you give appropriate credit and cite the associated project or paper. Commercial use is not permitted without prior written permission from the authors.