rafmacalaba/data-use-focus-metadata-sft

Passage metadata SFT

Separate ChatML configs for source attributes and usage metadata. One row per related, high-confidence mention with complete usage fields and data_type. Empty-output passages, not-related/unclear mentions, low-confidence decisions, and incomplete labels are excluded. Splits carry over from agents_v3_kept unless the builder was run with --split-ratio, which re-deals whole documents (never single mentions) across the three splits; either way document boundaries are checked for leakage and the build fails if a document straddles. This is agent-annotated metadata, not independent human gold.

English-only corpus: passages written in French or Spanish (~76 mentions in train, 15 in val, 12 in holdout) are excluded from every split; --keep-non-english retains them. Short table fragments are kept because they carry too little text to judge.

Known label drift: the train lane was annotated under a looser usage_action policy than the val/holdout lanes (source 65% vs 31%), so a model trained on it collapses onto source. Train usage rows are upsampled so no action exceeds ~40% of that split; every distinct annotated row is kept. Read usage_action scores as agreement with a mixed labelling policy, not ground truth.

Each task has 7,200 distinct mentions: train 5,040, val 1,080, holdout 1,080 (70/15/15). Those are mentions, not rows: the usage train file repeats rows to keep one action under 40% of the split, so its own counts are 7,586 / 1,080 / 1,080 and its row share is not the split ratio.

manifest.jsonl records excluded mention occurrences and reasons. split_stats.json records source hashes and counts.

synth/ holds the generated lane: attributes_authored.jsonl and usage_authored.jsonl, one ChatML row per synthetic source, each carrying its own split field in the same 70/15/15 deal, plus authored_stats.json with the leakage and phrase-sharing numbers. Those passages were written around label values chosen first, so a value is correct by construction rather than grounded in a real document - read them as generated training data, not as a second annotation. synth/*_synth.jsonl is the older 10-row batch labeled from real passages.

synth/authored_stats.json states that lane's own sharing: worst_shared_phrase_share is 2.73% of its eval rows carry the single 8-gram most repeated across the train/eval line, against 0.8% in the labeled lane, while anchor_shared is 0. That is a memorisation risk, not a label error: a model can learn a closing from the generated train rows and repeat it on the generated eval rows, so scores over synth/ eval rows are the weaker measure. The labeled val/holdout rows are the safer eval set.