Datasets:
Quick start
Install
Use Python 3.11 and FFmpeg.
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
Configure
Fill in the model and text Judge URLs, served model IDs and API keys in .env. The model endpoint accepts chat-completions requests containing images and text. Keys can be empty for your own unauthenticated local endpoints. For videos, apply for Ego4D access and fill EGO4D_ACCESS_KEY_ID and EGO4D_SECRET_ACCESS_KEY in .env. See Download Guide for an existing AWS profile or local source cache.
Evaluate
Run all 1,000 questions over 630 video episodes:
bash scripts/evaluate.sh
The first run prepares the 27.6 GB benchmark media tree and caches only the 76 required Ego4D originals (about 48.3 GB). The other seven video sources come from Hugging Face. Install FFmpeg and allow additional space for sampled frames. API requests use your configured model and Judge services.
Results are saved in runs/my-model/ (or your configured RUN_DIR):
predictions.jsonl Model answers and rationales
judgments.jsonl Answer and rationale assessments
scores.jsonl Question-level scores
scores.summary.json Raw / Verified capability and macro scores
Continue a run
Repeat the same command after interruption. Saved terminal records are reused, and each stage verifies that the inputs and settings match. Use a new directory for a different model or question population.
Temporary network errors are retried twice per inference request. After fixing an endpoint problem, retry saved execution errors explicitly:
RETRY_ERRORS=1 bash scripts/evaluate.sh
Inference prints progress and reports successful responses, format failures, endpoint errors and pending questions. A failed stage returns a nonzero exit code. See Troubleshooting for recovery steps.
Next: model integration, local video inference, data format, scoring.
Recompute paper scores
bash scripts/replay.sh
Checks 19,000 question-level scores and 190 table values from the released scoring factors. Output: runs/paper_replay.json.