---
license: apache-2.0
base_model: Qwen/Qwen3-VL-8B-Thinking
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - process-reward-model
  - visual-reasoning
  - multimodal
  - math
  - sft
---

# VRPRM-sft

`VRPRM-sft` is the merged supervised checkpoint of VRPRM, a visual reasoning
process reward model. It first generates a global reasoning trace and then emits
a binary `0/1` judgment for each candidate solution step.

## Training

- Base model: `Qwen/Qwen3-VL-8B-Thinking`
- Data: 1,378 multi-turn conversations / 5,592 confident step labels
- Loss weights: 0.5 for global thinking, 2.0 for step judgments
- Tuning: LoRA r=16, alpha=32, lr=1e-4
- Training: 3 epochs / 513 steps, followed by LoRA merge

## Evaluation

VisualProcessBench official no-reference protocol:

| Model | Overall step macro-F1 |
|---|---:|
| VRPRM-sft | 66.15 |

## Usage

Use the two-stage global-thinking protocol from the code repository. First ask
for one `<think>...</think>` block, then query each step with guided choice
`0/1` while feeding previous judgments. The evaluator is
`benchmarks/visualprocessbench/eval_vpb_global_think.py`.

## Limitations

This model is intended for research on process reward modeling and multimodal
reasoning. It is not a general assistant and should not be used for automated
grading of high-stakes work without human review.