The same finetune_en.json recipe re-run from seed 2 beats the published
checkpoint on 7 of 8 metrics, measured on data/eval_en.jsonl (1500
records, RTX 3060):
metric previous this adapter
overall acc 0.7633 0.8587
overall ECE 0.0609 0.0225
choice acc 0.8340 0.9480
score acc 0.7120 0.9100
noul acc 0.7440 0.7180 <- only regression
noul ECE 0.1011 0.0202
Adapter, choice head, per-primitive temperature, adapter config and model
card all come from checkpoints/en. The previous adapter stays reachable at
revision=bac4de4 and earlier ones in the history.
fix: resync jeba-en checkpoint and card with checkpoints/en (v0.3.0)
The Hub still carried the pre-release card ('weights and metrics are
pending the RTX 3060 training run') next to a v0.2.0 checkpoint whose
temperature_calibration.json reports ece_after 0.1406 / temperature
0.75 - none of which matches the numbers published in
benchmarks/report.md (English 0.763 / ECE 0.061).
Re-uploads the adapter, choice head, per-language temperature and
adapter config straight from checkpoints/en (the source of the
published metrics), plus the corrected model card.
Add dedicated choice head (v3) and refresh calibration
