data/README_D.md

# Tier D: Evaluation Data

This directory contains the exact QA evaluation splits consumed by the seven
reported downstream tasks, plus the local metric implementation used by the
evaluation scripts.

```text
qa_eval/textvqa/       validation split
qa_eval/chartqa/       test split
qa_eval/gqa/           testdev-balanced images and instructions
qa_eval/table3/        AI2D test, DocVQA validation, VQAv2 validation, MME test
third_party/lmms-eval/ VQA normalization/metric implementation
```

Unused training splits and alternative GQA/ChartQA exports are excluded. The
ImageNet-1K train/validation image folders are distributed separately in
sharded tar archives under `imagenet1k/` so that the repository does not need
to track approximately 1.3 million individual image files.

This repository is private. ImageNet and each QA dataset retain their original
licenses and terms; do not make these data artifacts public without confirming
redistribution rights.