PSALM Ego-Exo4D Exo2Ego Small (4 epochs)
Private experiment artifact for the PSALM Ego-Exo4D-trained ZSL baseline.
- Training split: first
floor(123381 / 3) = 41127released Exo2Ego FullTrain records, preserving source JSON order. - Training: 4 × H100, per-device batch 12, 4 epochs, 3424 optimizer steps.
- Source revision:
625a17fe9a37b5ac32ec5556f21646f9b6ed3ee2. - Data-list SHA-256:
86a01325b307be97c539815d3bd6de0cddebf3bc54d078f589a00858ceb40990. - Training launcher SHA-256:
2f1c8c76a352b4714b6191d4101995b5ae70eaa093c210680d55223c45124dc5.
The audit/ directory contains the launcher, environment/data receipts, source-path-only patch, full training log, and exit receipt. SHA256SUMS covers every uploaded artifact except itself.
This checkpoint is intended for the seven audited evaluations requested in the associated experiment: DROID and BehaviorSim in Ego→Exo, Exo→Ego, and Exo→Exo transfer settings, plus the EgoExo4D-selected Exo→Exo control set.