Datasets:

psychemme
/

PsycheMME

Apply through the official Ego4D page. Once approved, paste the supplied AWS Access Key ID and Secret Access Key into these fields in your local .env:

EGO4D_ACCESS_KEY_ID=
EGO4D_SECRET_ACCESS_KEY=

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.

Fill EGO4D_SESSION_TOKEN only if it was included in your official credentials. An existing named AWS profile can be used instead through EGO4D_PROFILE.

Prepare all videos

bash scripts/download.sh

The script downloads annotations and the other seven video sources from Hugging Face. For Ego4D it uses the official downloader to fetch only the 76 source videos needed for the benchmark, then prepares 142 clips at the recorded source offsets using the benchmark's video and loudness-processing recipe. All 630 clips are placed in data/PsycheMME/benchmark/media/video/ with their existing media IDs.

Allow approximately 76 GB for the 48.3 GB Ego4D source cache and 27.6 GB prepared benchmark media tree, plus temporary files and frame caches. Source videos are reused on subsequent runs. The official CLI handles its download cache; completed clips are checked and reused after interruption. Renew expired credentials through Ego4D and rerun the same command.

If you already downloaded the official Ego4D originals, set EGO4D_SOURCE_DIR to that download root or its v2/full_scale/ directory. Existing originals are used directly.

Preparation writes metadata/media.local.jsonl, the local video index used by scripts/evaluate.sh. Questions, references, ASR and the published metadata remain unchanged.

Evaluate

Configure the model and Judge fields in .env, then run:

bash scripts/evaluate.sh

This also prepares missing media, so the separate download command is optional.

For annotations only:

bash scripts/download.sh --metadata-only