ANNOUNCE box benchbox · card 0 (one RTX 3090, 250 W) · arm 11 · train encoder `longtable_voices` · scripts/train.py --task tasks/longtable_voices --epochs 6 --bs 8 · 2026-09-21T22:17:44Z longtable_voices heldout n=30 trained 23.3% openjev 86.7% gemma4 70.0% floor 16.7% 14s 598 MB ANNOUNCE box benchbox · card 0 (one RTX 3090, 250 W) · arm 11 · train encoder `doorman_gate` · scripts/train.py --task tasks/doorman_gate --epochs 6 --bs 8 · 2026-09-21T22:18:05Z doorman_gate heldout n=15 trained 73.3% openjev 100.0% gemma4 93.3% floor 73.3% 9s 598 MB ANNOUNCE box benchbox · card 0 (one RTX 3090, 250 W) · arm 11 · train encoder `fieldexam_grounded` · scripts/train.py --task tasks/fieldexam_grounded --epochs 6 --bs 8 · 2026-09-21T22:18:21Z fieldexam_grounded heldout n=12 trained 50.0% openjev 100.0% gemma4 100.0% floor 50.0% 9s 598 MB