Training from the archived inputs
open1-b7g-e4 derives from open1-b5-e4: the same rows in the same order, with labels.guides
relabelled from tool calls (training/turn-load/relabel_guides.py). Train the B7 and
B7G heads with training/turn-load/train.sh:
train.sh GPU B7 RUN:--families tools,guidesoncorpora/independent.json,--tool-truth consensus --tool-weight none --tool-negatives 24 --pos-weight 6 --lr 0.0005 --seed 11 --batch-size 64 --epochs 45 --patience 8.train.sh GPU B7G RUN: the same with--families guides.
Score a head on every tool and guide option with training/turn-load/turn_probe.py and
audit omissions with training/turn-load/guide_audit.py. The backbone is
convaiinnovations/laya-multilingual at revision
e4e9ddf21a7b1903b7acffd8814ad4307bf63a67; obtain that revision explicitly.