diff --git a/persistent-memory.md b/persistent-memory.md index 23121db..6d6cf8a 100644 --- a/persistent-memory.md +++ b/persistent-memory.md @@ -117,18 +117,31 @@ no longer deployed sidecars here. See Recent decisions.) _As of 2026-09-17 ~09:00 PT._ -### IN FLIGHT: lv-mccarthy D4 pairs, generating on pfi-gx10 +### DONE: lv-mccarthy D4 pairs built — NEXT is `train_pairs_lora.py` -`~/lv-mccarthy/run-pairs.sh` launched 2026-09-17 ~11:50 PT: **3,722 train** then **276 val** -pairs against the `gen` seat, ~25 min. Seat fingerprint `vllm-0.1.1.dev50+geed1f3d0c-601ad866` -— **identical to Hemingway's run**, so the two pair sets are generator-comparable. -Watch `pairs/train.log`; `pairs/.complete` is written at the end. +`~/lv-mccarthy/pairs/` on gx10, 2026-09-17 11:55 PT via `run-pairs.sh` (committed at +`scripts/mccarthy-corpus/run-pairs.sh`). -Then `train_pairs_lora.py`. Before the eval, honour D1's pre-registration: a -**punctuation-normalised secondary read** of the voice axis. +``` +train 3,673 pairs of 3,722 passages reject 1.32% 461,172 words 1,377 s +val 269 pairs of 276 passages reject 2.54% 33,477 words 98 s +``` -⚠ The corpus under it was **rebuilt at ~11:45 PT** — see the split-name leak below. Old tree -kept at `corpus-{clean,renamed}.pre-splitfix` on gx10. Recipe: `scripts/mccarthy-corpus/RUNBOOK.md`. +⭐ **The val split is the largest fixture in the line** — 269 against Hemingway's 200 and +Brontë's 44 in-band, the structural cause of Brontë's underpowered voice axis. +Generator fingerprint `vllm-0.1.1.dev50+geed1f3d0c-601ad866`, **identical to Hemingway's +run and unmoved mid-run**, so the two pair sets are generator-comparable. + +Verified in the produced pairs, not assumed: **0 mid-sentence line breaks** in all 3,942 +responses, **0 chapter-argument responses**, and **0.0 quote marks per 10k words** in both +splits — McCarthy's unquoted dialogue survives rename → chunk → reflow → pair intact. +`sourcename` rejected **14 beats** that named a character the rename removed; Hemingway's +shipped pairs had no such filter at all. `narrator_retries` 101 of 3,673 (2.8%) against +Hemingway's 8.7%, which is what third-person prose looks like. + +Then, before the eval, honour D1's pre-registration: a **punctuation-normalised secondary +read** of the voice axis — and note the `mccarthy` register now states the tics, so the +base control arm gets them too. ### OPERATOR CALL — lv-hemingway carries 3 of the same hidden survivors and is LIVE