[purinda]

Summary

Training data and pipeline changes for Deep's candidate E. E is not shipped and not installed. It is better than D on the four skills it targets and worse on every broad set (results below), so D remains the candidate.

New training families

Each family has a generator in scripts/e_seeds/ and a hand-checked gate set in Training/eval/.

• slot-corrections: a correction inside one slot keeps the words before the slot. "The cost of parking, no wait, petrol has doubled" becomes "The cost of petrol has doubled.", not "Petrol has doubled.".
• spoken-numbers: numbers are kept exactly as said. The number rules (Write spoken numbers in digits after cleanup, from Medium up #55) write digits after the model.
• marker-lists: steps counted out loud in mixed styles are kept as said, so the list rules can number them.
• Word fragments and whole-word controls, as transcript seeds: "con consider" against "an animal", "new newsletter".

Pipeline

• prepare --extra-dir DIR adds a family folder to Deep's seeds, leaving out rows a split already holds. It reproduces E's prepared seeds row for row.
• merge-asr reconciles numbers. A number said in parts now reconciles with the single figure the recognizer writes ("seven point six" → 7.6, "zero four four six" → 0446). Before, the audit quarantined most spoken-number rows.
• Two category fixes: whole-word controls are filed as unchanged, and counted steps as series. Before this, the audit quarantined every one of them without failing the run.
• A moved-stops input variant: a comma or full stop a word away from where it belongs.
• The scorer now accepts either case for the word after a placeholder that opens a line. This doesn't change Train Deep on measured speech-to-text inputs, punctuated every way #51's numbers.

The branch merges the word-fragment check (#52), which E's targets need.

Candidate E

Trained from the base model on 11,440 train, 1,182 valid and 1,158 test rows, for 1,859 iterations. The best validation loss was 0.0152, at iteration 1,400. The table counts cases right by meaning and required punctuation; fallbacks are in brackets.

Set	Cases	1.1.0	D	E	E vs D
Gate: corrections inside a clause	41	36	36	39	+3 −0
Gate: word fragments and controls	31	27	26 (1)	29	+3 −0
Gate: spoken numbers kept exactly	42	36	35	38	+3 −0
Gate: list markers, placeholders	26	18	19	25	+7 −1
E's test split	1158	1019 (17)	1077 (10)	1073 (7)	+29 −33
Held-out measured pairs	837	772 (4)	799 (1)	783 (4)	+9 −25
D's test split	939	832 (9)	877 (8)	860 (8)	+14 −31
Held-out pairs as other recognizers write them	714	483 (3)	660 (4)	634 (9)	+18 −44
Hand-written eval	343	314 (3)	310 (4)	299 (2)	+7 −18
Standard's correction set	515	498	500	491 (4)	+5 −14

Where E loses. Most losses are text E leaves as said where D fixes it:

• garbled corrections ("the after tomorrow");
• grammar ("Uma were late");
• lists and letters it no longer lays out.

The new families are mostly keep-as-said: whole-word controls, numbers kept exactly, and counted steps kept as text in fields that take several lines. That probably made the adapter more conservative. A next candidate should rebalance the families against layout and grammar, rather than add more of them.

Test plan

• prepare-cleanup-data: 70 Python tests. cleanup-scoring: 6.
• The generators' --check: no raw text is shared with any held-out set. Deep's check accepts every target.
• Full pipeline for E: speech, recognition, merge, audit, quarantine, training, and measurement of E, D and 1.1.0 with the same build.
• Swift Cleanup, CleanupTraining, Styles, Dictation, SpokenCommands and Shared tests pass on this branch merged with Show Medium's fallback at Deep only if Deep's check accepts it too #53 and Write a name before a mail provider's domain as an email, and lay out lists that mix marker styles #56. Rust lt-cleanup, lt-styles, lt-dictation, lt-spoken-commands and lt-shared tests pass.

🤖 Generated with Claude Code

[purinda]

Superseded by #58, which merged this work into main with the other changes behind candidate E. It shipped in 1.2.0. (GitHub closed this PR when its base branch was deleted.)