form-autofill-head

Fusion head for the Firefox triple-encoder autofill field classifier. Shipped as ONNX: onnx/model.onnx (fp32) and onnx/model_quantized.onnx (int8, per-channel + reduce_range). Must be paired with the matching Mozilla/form-autofill-embed encoder — the fusion weights are tied to that encoder's embedding space, so update both together.

• Input fused: float32 [batch, 1920] = concat[e_cur, e_prev, e_next, e_cur-e_prev, e_cur-e_next] (each 384-d, mean-pooled from Mozilla/form-autofill-embed).
• Graph: Linear(1920->384) -> GELU(erf) -> Linear(384->66).
• Output logits: [batch, 66] over autofill field types (see config.json id2label).

Provenance

• Run: argo-autofillflow-kbns9 (triple, encoder_layers 4, head_interactions, gen_to_real_ratio 8; namespace production:autofillflow-0-egrc).
• Training data: -relabel — o4-mini-audited relabel of the shipped corpus + a fresh multilingual Common-Crawl batch (CC-MAIN-2026-34: en/de/es/ja/it/nl/pt); select_option + input_attributes features.
• Base encoder: rolf-mozilla/minilm-pruned-notest-relabel (vocab 18,473, 4 layers).
• Metrics (full triple, test): total 0.9103 / close 0.9266; name P0.841 / R0.768. int8 encoder lossless (fp32==int8 total/close).

int8 weights verified within ±64 (safe for the AVX2 int16 accumulator, Bugzilla 2064781).