license: apache-2.0
library_name: transformers
pipeline_tag: video-text-to-text
DAR-R1
This is the official model checkpoint for DAR-R1 (built on top of Qwen2.5-VL-3B-Instruct), presented in the paper: Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset.
Overview
Dynamic Affective Reasoning (DAR) is a viewer-centric video emotion benchmark. Instead of assigning a single static label to a whole clip, DAR asks a model to identify when the viewer's emotion changes (affective segmentation), what the fine-grained emotion is (fine-grained emotion classification), and why the visual event triggers that affective reaction (affective reasoning).
The benchmark contains 15,087 videos, 36,908 event-aligned affective segments, and 27 emotion categories. Each segment includes a temporal span, an emotion label, and a visually grounded causal rationale.
DAR-R1 is trained using a two-stage framework:
- Cold-Start SFT: Adapts Qwen2.5-VL-3B-Instruct to the structured DAR output format.
- GRPO Training: Uses Group Relative Policy Optimization (GRPO) to refine temporal localization, emotion prediction, and reasoning quality.
Resources
- GitHub Repository: Zhang-Zhiyan/DAR
- Hugging Face Dataset: aiaiaizzy/DAR-R1
- Paper: arXiv:2607.10238
Quick Start (Evaluation)
To run evaluation using the official script from the repository:
python test.py \
--model-path /path/to/DAR-R1 \
--test-jsonl /path/to/DAR/test.jsonl \
--video-root /path/to/DAR/videos \
--output-jsonl /path/to/outputs/dar_r1_test_predictions.jsonl \
--batch-size 8
Citation
If you find this model or the DAR benchmark useful in your research, please consider citing:
@misc{zhang2026benchmarkingdynamicaffectivereasoning,
title={Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset},
author={Zhiyan Zhang and Peipei Song and Jinpeng Hu and Jingyang Jia and Xun Yang and Xiaojun Chang},
year={2026},
eprint={2607.10238},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.10238},
}