Mozilla/smart-form-fill-relevant-tabs

Smart Form Fill - Relevant Tabs (Call 2 / select_tabs) eval (v2)

Given the form the user is filling (the anchor) and a set of open browser tabs (candidates, title + url only), pick which tabs are relevant sources for filling the form. Each row is one session.

Model input (per row)

page{anchor_title, anchor_url} + anchor_fields + session_tabs (each {id,title,url}, metadata only). The model returns a relevance judgement per tab id.

Labels

relevance = score(anchor.field_types INTERSECT candidate.provides); distinctive types (passport-/cc-/ssn/linkedin/github/work-authorization/id-number/nationality/referral-source) x1.5, common x1, other/contextual 0; relevant when score >= 3.0. Candidate provides: a form provides its own fields; a content-source tab (LinkedIn/GitHub/resume/account/ order/flight/SSA/wallet/...) provides a curated set; a distractor (non-form page) provides nothing. Labels are computed from provides (never shown to the model).

Columns

• —input: anchor_title, anchor_url, anchor_fields (JSON), session_tabs (JSON [{id,title,url}])
• —labels: relevant_tab_ids (JSON, score>=3), tab_relevance (JSON {id:score}, for NDCG/top-N)
• —slicing: session_type (hasrelevant/nullcase), session_size (n = #tabs, in {6,12,20,30,40,50}), anchor_dataset, anchor_distinctive_types, anchor_field_types, tab_kinds (JSON {id:form|content_source|distractor})
• —convenience: num_relevant, tabs_debug

Notes

• —n (session size) is a dataset axis (getopentabs product cap is 30; 40/50 stress-test beyond it) -> slice metrics by session_size.
• —m (maxSelectedTabs, "limited by pageExtractor") is a scoring-time sweep -> the row holds the full ranking.
• —379 sessions (290 hasrelevant / 89 nullcase; 75 dense with 5+ relevant). Realistic composition: 1-2 true relevant tabs per session (5-8 in the dense slice) amid mostly content/distractor tabs. Relevance = category match (commerce/job/travel/identity source), validated 11/11 vs human labels. English-text only. Synthetic proxy.

Load

from datasets import load_dataset
ds = load_dataset("Mozilla/smart-form-fill-relevant-tabs", split="train")