The card published a per-language table, 21 of 26 rows at exactly 1.000, and no per-label figures, while listing eight label types. The language axis cannot show which label was assigned, and label assignment is where this model fails.
The per-entity table already existed in the run's heldout_ner_eval.json and was never published. LOCATION has zero gold spans and has therefore never been scored. NATIONAL_ID recalls 0.0769 with 151 of 208 spans mislabelled. CARD puts a spurious span on 192 of 520 rows. Leaked tokens are 0 on every row, so these are wrong names on covered spans rather than disclosures.
Also replaces a literal '? of 300 decisions differ' left by a generator field that never filled; the export manifest says 0 of 300.
Upload eval/piiguard_eval.json with huggingface_hub
Deletes the superseded INT8 and fp32 ONNX exports, the torch training args, and metrics.json, whose test_f1 of 1.0 was measured on a split drawn from the same templates as the training half. Replaces model.safetensors with the adopted run's weights. The card carries the held-out numbers, which are the ones that describe behaviour on text the generator did not write.
