photonmz/bwm-sft-wa

Provenance

Every row carries Tülu-style lineage:

• —content_hash — sha256 over (system, user, assistant, site, is_noop, sha256(image bytes)); stable content identity (oversampled duplicate rows share it).
• —source — <collection>/<site>, collection ∈ {webarena, liveweb}, inferred per trajectory (a trajectory touching the WebArena instance host is webarena).

Decontamination (report-only; no rows removed)

Checked at revision 320093e6918a84285cc9e2e8e1b9aee3a5ea89e0: 8-gram word overlap of train prompts vs valin/valout prompts (GOAL/CURRENT URL/ACTION lines only), plus val_out site- and goal-template-identity (docs/plan/05-evals.md).

• —removal rate (flagged, not removed): 95.7% (605/632 train rows)
• —breakdown: ngram∩valin 600, ngram∩valout 0, exact prompt 0, valout site 0, valout template family 18
• —trajectories straddling train/val_out (site-wise split cuts multi-site trajectories row-wise): 0

val_in is an in-template dev split, not OOD. Train and valin share the single WebArena instance host and its small intent-template pool, so valin measures in-template generalization only. Never present val_in metrics as out-of-distribution results; the 95.7% n-gram flag rate above is structural URL/template overlap, not row duplication.

Changelog

• —2026-06-12 (rev fb179751004b308465d0ef8e1317e1ed4899f5b6): add content_hash + source provenance columns and record decontamination stats; data otherwise unchanged from 320093e6918a84285cc9e2e8e1b9aee3a5ea89e0.
• —2026-06-12 (rev d7d8cd31c49d30e5b6d7fa431ec40852fc2046d2): decontamination removal per PR #12 review sign-off — drop the 5 train/valout-straddling trajectories' rows from valout (48 → 33). train and val_in unchanged.