---
license: mit
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-2B-Instruct
tags:
- image-quality-assessment
- iqa
- qwen3-vl
- mr-iqa
- grpo
---

# MR-IQA 2B

This repository contains the MR-IQA 2B no-thinking model from the original 2026-06-17 training run.

## Model

- Base model: Qwen3-VL-2B-Instruct
- Training run: `visualquality_r1_non_thinking_2b_full_zero3_dmc_rawmargin_l2_noordinal_2b_20260617_1257`
- Published checkpoint: `best_model`
- Best validation epoch: 3
- Validation set: KONIQ 200-sample validation split with dataset seed 42
- Validation result: SRCC 0.9317891406, PLCC approximately 0.94

## Files

This upload keeps only inference-required resources:

- `model.safetensors`
- `config.json`
- `generation_config.json`
- `processor_config.json`
- `tokenizer.json`
- `tokenizer_config.json`
- `chat_template.jinja`

Training-only artifacts such as optimizer states, scheduler states, RNG states, trainer state, DeepSpeed ZeRO folders, and local path metadata are intentionally excluded.

## Prompt Format

The training/evaluation prompt asks for an overall image quality score as a float between 1 and 5, rounded to two decimal places, with the final answer in `<answer>...</answer>` tags.

## Path Hygiene

The uploaded model files do not require local training paths. Provide image files through your own inference or evaluation pipeline.