QGO-8B / README.md
PM4Bench-Release
Pin paper references to arXiv v3
a885f4e
|
Raw History Blame Contribute Delete
4.44 kB
---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-8B-Thinking
tags:
- qwen3-vl
- vision-language
- multilingual
- ocr
- grpo
datasets:
- DatasetMan/PM4Bench-QGO-Train
---
<div align="center">
# QGO-8B
**OCR-centric reinforcement learning for multilingual LVLMs**
[![arXiv v3](https://img.shields.io/badge/arXiv-v3-B31B1B.svg?logo=arxiv&logoColor=white)](https://little-bird-vodka.github.io/PM4Bench/paper-v3.html)
[![Project Page](https://img.shields.io/badge/Project-Page-145C73.svg?logo=githubpages&logoColor=white)](https://little-bird-vodka.github.io/PM4Bench/)
[![Code](https://img.shields.io/badge/GitHub-Code-181717.svg?logo=github&logoColor=white)](https://github.com/opendatalab/PM4Bench)
[![Benchmark](https://img.shields.io/badge/Hugging_Face-Benchmark-FFD21E.svg)](https://huggingface.co/datasets/songjhPKU/PM4Bench)
[![Training Data](https://img.shields.io/badge/Hugging_Face-QGO_Train-FFD21E.svg)](https://huggingface.co/datasets/DatasetMan/PM4Bench-QGO-Train)
</div>
## Overview
The paper [*Benchmarking and Boosting Multilingual Capabilities of LVLMs via
OCR-Centric Reinforcement Learning*](https://little-bird-vodka.github.io/PM4Bench/paper-v3.html)
introduces the strictly parallel PM4Bench benchmark and finds that OCR is a key
source of cross-lingual performance gaps under vision-only input. QGO trains
on fully synthesized OCR data with a dual reward for recognition accuracy and
reasoning-preserving response structure. It improves average score on every
PM4Bench task and setting and transfers gains to external multilingual OCR
benchmarks.
QGO-8B is the OCR-centric GRPO checkpoint derived from
[`Qwen/Qwen3-VL-8B-Thinking`](https://huggingface.co/Qwen/Qwen3-VL-8B-Thinking).
This repository contains the complete global-step-200 BF16 checkpoint used in
the PM4Bench paper.
## Model details
- Architecture: `Qwen3VLForConditionalGeneration`
- Base model: `Qwen/Qwen3-VL-8B-Thinking`
- Training method: GRPO
- Released checkpoint: global step 200
- Weight format: safetensors, four BF16 shards
- Indexed tensor bytes: 17,534,247,392
- License: Apache-2.0
- Tested Transformers version: 4.57.6
## Usage
```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
model_id = "DatasetMan/QGO-8B"
model = Qwen3VLForConditionalGeneration.from_pretrained(
model_id,
dtype="auto",
device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)
```
The full BF16 weights are approximately 17.5 GB before runtime allocations.
Plan GPU/CPU memory for weights, vision inputs, KV cache, and generation in
addition to the checkpoint size.
Use the base model's official chat template and greedy decoding for PM4Bench
evaluation. Task prompts and evaluation code are provided in
<https://github.com/opendatalab/PM4Bench>.
## Training
The released data is available at
<https://huggingface.co/datasets/DatasetMan/PM4Bench-QGO-Train>. The recipe
uses 32 prompts and 8 rollouts per prompt (256 trajectories per step), AdamW
with learning rate `1e-6`, BF16, and eight GPUs.
## PM4Bench evaluation
| Model | MDUR trad. | MDUR vision | MIQA trad. | MIQA vision | MSOCR | MGUI |
|---|---:|---:|---:|---:|---:|---:|
| Qwen3-VL-8B-Thinking | 38.55 | 34.88 | 53.63 | 47.69 | 1.53 | 78.30 |
| QGO-8B | 46.82 | 40.84 | 55.24 | 51.06 | 8.17 | 80.00 |
These are the audited paper results. MDUR and MGUI are percentages, MIQA is
the six-dimension judge score on a 10-100 scale, and MSOCR is on a 0-40 scale.
## Limitations
QGO-8B targets multilingual OCR robustness. It inherits limitations and risks
from the Qwen base model and is not guaranteed to improve every downstream
task or language. Coordinate outputs, OCR transcriptions, and long-form
reasoning should be validated before use in consequential applications.
## Cite Us
<pre><code class="language-bibtex">
@misc{gao2026benchmarkingboostingmultilingualcapabilities,
title={Benchmarking and Boosting Multilingual Capabilities of LVLMs via OCR-Centric Reinforcement Learning},
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},
year={2026},
eprint={2503&#46;18484},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv&#46;org/abs/2503&#46;18484v3},
}
</code></pre>