Mozilla/minilm-multi-address-autofill

tinybert/minilm address-autofill — bb-intl (relabel)

Monolithic (single-sequence, bb context) field-type classifier for Firefox form autofill. Predicts an autocomplete-style token per form field from its mlData (id/name/placeholder/label + bb/aa neighbor context).

Provenance

• Architecture: bb (monolithic single-sequence classifier)
• Base encoder: rolf-mozilla/minilm-pruned-notest-relabel — multilingual MiniLM-L12 pruned to a 18,473-token vocab (train+val only, test excluded), 4 layers.
• Training data: -relabel — the shipped corpus re-labeled with 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) merged into data/common_crawl.
• mlData features: select_option (option-range token) + input_attributes (**maxlen<N> / **inputmode).
• Metaflow run: argo-autofillflow-8wpwd (namespace production:autofillflow-0-egrc)
• Hyperparams: encoder_layers 4, 6 epochs, batch 32, lr 7e-5, warmup 0.1, weight_decay 0.05, gen_to_real_ratio 2.0, subsample_validation.

Metrics

split	total acc	close acc	name recall
test (data/testing)	0.8873	0.9117	0.826
validation	0.9042	0.9270	0.940

name recall improved ~+16 pt over the previous bb baseline (0.783 → 0.940 on validation) — the fix for full-name fields mislabeled as family-name.

Files

• model.safetensors, config.json, tokenizer.json — fp32 PyTorch model.
• onnx/model.onnx — fp32 ONNX.
• onnx/model_quantized.onnx (= model_int8.onnx) — int8, per-channel + reduce_range (weights within ±64, safe for the AVX2 int16 accumulator — Bugzilla 2064781). 14.7 MB.

int8 accuracy (lossless)

fp32 0.8873 → int8 0.8861 on testing-relabel.txt (Δ −0.12 pt).