form-autofill-embed

Quantized (int8) shared encoder for the Firefox triple-encoder autofill field classifier. Vocab-pruned multilingual MiniLM (BERT arch, 4 layers, hidden 384, vocab 18,473). Use as a feature-extraction model and mean-pool the last_hidden_state to a 384-d per-field embedding. Weights: onnx/model.onnx (fp32) and onnx/model_quantized.onnx (int8, per-channel + reduce_range). Pairs with Mozilla/form-autofill-head — the encoder and head must be updated together (the head's fusion weights are tied to this encoder's embedding space).

Provenance

• Base: multilingual MiniLM-L12 pruned to 18,473 tokens on the relabel train+val (test excluded), 4 layers.
• Training data: -relabel — the shipped corpus re-labeled via an o4-mini audit (name/label corrections) plus a fresh multilingual Common-Crawl batch (CC-MAIN-2026-34: en/de/es/ja/it/nl/pt), with select_option + input_attributes mlData features.
• Run: argo-autofillflow-kbns9 (triple, encoder_layers 4, head_interactions, gen_to_real_ratio 8; namespace production:autofillflow-0-egrc).
• 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).