teenyfill-149m

TeenyFill 149M — local web-form autofill

A 149,605,633-parameter ModernBERT cross-encoder that fills web forms from your own data, entirely on-device. Score every (field, value) candidate action on a page; apply the confident ones; never submit. Ship it inside test automation, RPA flows, or privacy-sensitive autofill.

Use through the teenyfill Python package (recommended):

from teenyfill.autofill import FormFiller
filler = FormFiller.from_pretrained("harrrshall/teenyfill-149m")
report = filler.fill(page, {"Full name": "Ada Lovelace", "Email": "[email protected]"})

Measured performance (all receipted in the training repo)

• Frozen Mind2Web 8-website holdout (never trained on): 27/57 field-exact fills — equal to a 32B teacher under the same protocol, at 1/200th the size.
• FormFactory official splits (product API, end-to-end): A11 8/8, E12 7/8, F11 8/8 form-exact (151/152 fields).
• Live websites (real network, real JS): demoqa React form 5/5 fields, practicesoftwaretesting 4/4, herokuapp login 2/2; multi-step flows advance through Next/Continue and stop before final submission by default.
• Latency: ~0.04 s/page on an A100, ~0.5 s/page on CPU.

Limitations

Native HTML controls only (no custom div dropdowns/canvas/CAPTCHA); no shadow DOM/iframe traversal yet; English-first; adjacent same-type fields with ambiguous one-word values (first/last name in label-less table layouts) can confuse it — review the fill report before submitting anything that matters.

Training

Fine-tuned from ModernBERT-base on gold form-filling actions from Mind2Web (CC-BY-4.0) and FormFactory (MIT) plus synthetic forms; listwise softmax-CE with mined hard negatives. No third-party model outputs were used. Full audit trail (including negative results) in the project repository.