---
pretty_name: 'ocrbench: 500-sample lite evaluation'
task_categories:
- visual-question-answering
size_categories:
- n<1K
tags:
- benchmark
- lite
- reproducible-subset
configs:
- config_name: default
  data_files:
  - split: test
    path: data/ocrbench/*.parquet
---

# ocrbench: 500-sample lite benchmark

500 evaluation examples sampled with seed **42**. Sampling: uniform random rows without replacement.

## Sources and attribution

- [echo840/OCRBench](https://huggingface.co/datasets/echo840/OCRBench/tree/92a54bd1384387c178d5a07140a2d85e0a3d12e1), revision `92a54bd1384387c178d5a07140a2d85e0a3d12e1`.

Original dataset and source-image terms apply; this subset does not relicense them. See the linked upstream datasets for their license and citation information.

## Reproducibility

`subset.json` records the pinned sources, selected source indices, decoded-image/label hashes, native task configurations, seed, and sampling procedure. `_lite_source_index` and `_lite_task_id` identify each example in its prepared source split. Image bytes are embedded in the Parquet files; no cluster paths are required. Original pixels, questions, reference labels, and scoring fields are preserved.

This is an evaluation subset. Its scores are not full-benchmark scores. Use the native benchmark metrics from [foveated_vlm](https://github.com/Gabesarch/foveated_vlm). This benchmark uses automatic scoring.