Datasets:
PM4Bench-Release commited on
Commit ·
9cd25ff
1
Parent(s): e5aaf81
Update paper links, resource navigation, and citation
Browse files
README.md
CHANGED
|
@@ -27,9 +27,33 @@ configs:
|
|
| 27 |
path: data/test/*.parquet
|
| 28 |
---
|
| 29 |
|
|
|
|
|
|
|
| 30 |
# PM4Bench QGO training data
|
| 31 |
|
| 32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
- 19,500 training rows
|
| 35 |
- 500 validation rows
|
|
@@ -47,5 +71,19 @@ The dataset is synthetic and released under Apache-2.0.
|
|
| 47 |
source digest, release digest, and the only permitted transformation. The
|
| 48 |
repository-level `SHA256SUMS` covers the complete public snapshot.
|
| 49 |
|
| 50 |
-
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
path: data/test/*.parquet
|
| 28 |
---
|
| 29 |
|
| 30 |
+
<div align="center">
|
| 31 |
+
|
| 32 |
# PM4Bench QGO training data
|
| 33 |
|
| 34 |
+
**Synthetic multilingual OCR data for QGO reinforcement learning**
|
| 35 |
+
|
| 36 |
+
[](https://arxiv.org/abs/2503.18484v3)
|
| 37 |
+
[](https://little-bird-vodka.github.io/PM4Bench/)
|
| 38 |
+
[](https://github.com/opendatalab/PM4Bench)
|
| 39 |
+
[](https://huggingface.co/datasets/songjhPKU/PM4Bench)
|
| 40 |
+
[](https://huggingface.co/DatasetMan/QGO-8B)
|
| 41 |
+
|
| 42 |
+
</div>
|
| 43 |
+
|
| 44 |
+
## Overview
|
| 45 |
+
|
| 46 |
+
The paper [*Benchmarking and Boosting Multilingual Capabilities of LVLMs via
|
| 47 |
+
OCR-Centric Reinforcement Learning*](https://arxiv.org/abs/2503.18484v3)
|
| 48 |
+
uses PM4Bench to show that OCR is a key source of cross-lingual performance
|
| 49 |
+
gaps when text is rendered visually. QGO addresses that finding with GRPO on
|
| 50 |
+
synthetic OCR data, without task-specific VQA or GUI supervision. This
|
| 51 |
+
repository contains the exact training and validation corpus used for the
|
| 52 |
+
released [QGO-8B checkpoint](https://huggingface.co/DatasetMan/QGO-8B).
|
| 53 |
+
|
| 54 |
+
## Dataset contents
|
| 55 |
+
|
| 56 |
+
The synthetic multilingual OCR corpus contains:
|
| 57 |
|
| 58 |
- 19,500 training rows
|
| 59 |
- 500 validation rows
|
|
|
|
| 71 |
source digest, release digest, and the only permitted transformation. The
|
| 72 |
repository-level `SHA256SUMS` covers the complete public snapshot.
|
| 73 |
|
| 74 |
+
The [QGO training recipe](https://github.com/opendatalab/PM4Bench/tree/main/recipes/qgo)
|
| 75 |
+
includes the launcher and exact reward implementation.
|
| 76 |
+
|
| 77 |
+
## Cite Us
|
| 78 |
+
|
| 79 |
+
```bibtex
|
| 80 |
+
@misc{gao2026benchmarkingboostingmultilingualcapabilities,
|
| 81 |
+
title={Benchmarking and Boosting Multilingual Capabilities of LVLMs via OCR-Centric Reinforcement Learning},
|
| 82 |
+
author={Junyuan Gao and Jiahe Song and Jiang Wu and Runchuan Zhu and Guanlin Shen and Shasha Wang and Xingjian Wei and Haote Yang and Weijia Li and Bin Wang and Lijun Wu and Conghui He},
|
| 83 |
+
year={2026},
|
| 84 |
+
eprint={2503.18484},
|
| 85 |
+
archivePrefix={arXiv},
|
| 86 |
+
primaryClass={cs.CV},
|
| 87 |
+
url={https://arxiv.org/abs/2503.18484},
|
| 88 |
+
}
|
| 89 |
+
```
|