Run on Your Video
Install FFmpeg and the evaluation package with a PyTorch CUDA build appropriate for your GPU:
python -m pip install "psychemme-eval[local-model] @ git+https://github.com/psychemme/PsycheMME.git@main"psychemme doctor --mode localpsychemme video --video /path/to/clip.mp4 \ --question "What social cues indicate that the person needs support?" \ --adapter psychemme/PsycheMME-Qwen3.5-4B-SFT \ --adapter-revision v1.0.2 --output runs/video.json
Add --asr transcript.txt to supply a UTF-8 transcript. The output records the response, sampled frame count, model revisions and finish reason.
The loader uses Qwen3_5ForConditionalGeneration, AutoProcessor and PEFT. Both PsycheMME adapters load directly on the original Qwen3.5-4B base.
For a local copy of this repository:
python infer_video.py --video /path/to/clip.mp4 \ --question "What social cues indicate that the person needs support?" \ --output runs/video.json
load_adapter.py provides the standalone loading example:
from load_adapter import load_adapterprocessor, model = load_adapter()
Benchmark Evaluation
Use the evaluation quick start for Infer → Judge → Score on the benchmark. The video command above answers a free-form question; benchmark evaluation uses the structured answer and rationale format.
Table with columns: Paper evaluation, Raw, Reasoning-Verified| Paper evaluation | Raw | Reasoning-Verified |
|---|
| Qwen3.5-4B + SFT | 72.5835 | 52.1345 |
Evaluation uses full-timeline 1 fps sampling capped at 64 frames, BF16/SDPA, batch size 1, greedy decoding, seed 42 and a maximum of 768 new tokens. Inputs include the question, supplied context, ASR and sampled frames. Training uses 32 frames. The metric averages across all eight capabilities over the 1,000-question population.
Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. See paper_evaluation_config.json for package versions and settings, and the reproduction guide for experiment details.
License
The model and adapter are distributed under Apache-2.0. Benchmark source terms are provided with the dataset.
Citation
@misc{hu2026psychemme, title={PsycheMME: A Psychology-Informed Benchmark for Social Understanding and Interaction in Videos}, author={Heran Hu and Yuhang Dong and Zhucun Xue and Kaiwei Chen and Yong Liu and Jiangning Zhang}, year={2026}, eprint={2610.28625}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2610.28625}}