PsycheMME

A Psychology-Informed Benchmark for Social Understanding and Interaction in Videos

Project · Code · Models · Evaluation Records · Training Data

Quick Start

For full video preparation, first apply for Ego4D access. Approval may take time, so request access first. You can load the annotations below while waiting.

Allow approximately 76 GB for the Ego4D source cache (48.3 GB) and prepared benchmark media (27.6 GB), plus temporary files and frame caches.

pip install datasets huggingface_hub

from datasets import load_dataset

questions = load_dataset("psychemme/PsycheMME", "questions", revision="v1.0.2", split="test")
asr = load_dataset("psychemme/PsycheMME", "asr", revision="v1.0.2", split="test")

Use Python 3.11 and FFmpeg. Prepare all benchmark videos with the evaluation repository:

git clone https://github.com/psychemme/PsycheMME.git
cd PsycheMME
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install '.[data]'
cp .env.example .env
# Fill your official Ego4D AWS Access Key ID and Secret Access Key.
bash scripts/download.sh

The script fetches only the 76 required Ego4D source videos from the official service and prepares 142 clips; the other seven sources (488 clips) are downloaded from this repository. See the Download Guide for local source reuse and storage.

Configure the supplied key pair in your local .env. The script passes it only to the official Ego4D downloader through a temporary local AWS profile, which is deleted after use. PsycheMME does not log the keys or upload them to GitHub or Hugging Face.

Benchmark

Content	Count	Path
Questions	1,000	benchmark/public/test.jsonl
Reference answers and evidence	1,000	benchmark/ground_truth.jsonl
Video episodes	630	benchmark/media/video/
ASR transcripts	630	benchmark/asr.jsonl
Episode mappings	630	metadata/episodes.jsonl
Source links and crop intervals	630	metadata/sources.csv

References are public. The evaluated model receives the question, options, supplied memory, video and ASR; reference answers and evidence enter scoring only.

memory_context and memory_dependency are optional and usually null: no separate history or dependency annotation is attached to those questions. Explicit history is provided for 40 questions, and dependency annotations for 54. Video and ASR remain part of the input.

Videos and Sources

The 630 videos occupy about 27.6 GB, including 61 clips at 4K UHD. Dimensions, frame rates, durations and audio information are listed in metadata/media.jsonl.

metadata/sources.csv provides original links, video/file identifiers, and source start/end times for all 630 clips. Source coordinates are distinct from clip-relative ASR and evidence timestamps. YouTube entries point to the original watch page; other entries point to source files or official dataset access pages.

Source Licenses

Source	License
MERCI+	Apache-2.0
CASTLE	CC BY-NC-SA 4.0
EasyCom	CC BY-NC 4.0
IndEgo	CC BY 4.0
Ego4D	Ego4D License Agreement
EgoCom	Dataset-specific terms
YouTube Vlog	Video-specific license; source links
YouTube Interview	Video-specific license; source links
PsycheMME numeric results and source-index metadata	CC BY 4.0

License references are collected in Source Licenses.

Evaluate Your Model

Install the evaluation package, configure your model and Judge endpoints and official Ego4D credentials in .env, then run:

bash scripts/evaluate.sh

Repeat the same command to continue an interrupted evaluation. See Quick Start for setup and outputs, and Data Format for fields.

Paper scoring factors and tables are available in PsycheMME-logs.

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}
}