---
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
  - reinforcement-learning
---

# VRPRM-rl

`VRPRM-rl` is the merged process-supervised RL checkpoint released for VRPRM.
It is initialized from `VRPRM-sft` and trained with GSPO-token/GRPO LoRA RL on
40K confident VisualPRM400K-derived examples.

## Training

- Initialization: `VRPRM-sft`
- Data: 40K train / 2K validation
- Rollouts: 16 per prompt, temperature 0.8, top-p 0.95
- Reward: `0.92 * step + 0.05 * format + 0.03 * think`
- Optimization: lr 1e-6, KL coefficient 0.2, 600 steps
- Public artifact: end-of-run merged checkpoint

## Evaluation

VisualProcessBench official no-reference protocol:

| Model | Overall step macro-F1 |
|---|---:|
| VRPRM-rl | 69.18 |

The public artifact is the end-of-run checkpoint. The run's highest RL
validation reward occurred earlier; comparisons should use the exact released
artifact and protocol documented in the code repository.

## Usage

Use the same global-thinking stepwise protocol as `VRPRM-sft`. The model is
evaluated with `benchmarks/visualprocessbench/eval_vpb_global_think.py`.

## Limitations

This research model is not a general assistant and is not suitable for
high-stakes automated grading without human oversight.