---
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
---

# QGO-8B

QGO-8B is an 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.