vh
61840f3131
lv-bronte is live on vllm-voices (fv-ml1 GPU0 :8027) alongside voices-base and lv-yarros. The seat lists all three; container healthy; GPU0 96092 -> 96090 MiB, so the adapter cost nothing measurable. IT DID NOT PASS ITS VOICE GATE, and the artifact says so in three places — this commit, a comment in the compose file, and a README beside the adapter on NFS — because an adapter found without its provenance will otherwise be read as a pass. VOICE FAIL +0.193 delta_cb vs base, against a 0.251 measured noise floor NOT COPIED PASS 8-gram hit-rate 0.00, longest 0 - identical to the control NO DAMAGE PASS ran-on +0.15 against a 0.400 floor Shipped on three grounds, none of them that the number was nearly good enough: it is additive (a named LoRA nobody reaches without asking for it), reversible (one compose line; hot-unload measures 0.003 s), and clean on the axis that carries actual risk - verbatim regurgitation of the source, on a public-domain corpus, measured against a positive control that saturates at 160. The voice result is UNDERPOWERED rather than absent: it closed 48% of the span from base to the same-author target and beat the control on every individual seed. The cause is structural - 81 val pairs against Hemingway's 200, from a 678k-word corpus against 994k - and neither more beats nor more seeds fixes it, because the floor is a range statistic and ranges widen with n. ckpt475 over ckpt925: indistinguishable on voice (0.017 apart), but ckpt925 has a verbatim 8-gram hit where this has none, and is 2.7x less stable seed-to-seed (0.251 vs 0.092) with a degeneracy probe showing no collapse to explain it.
88 lines
5.3 KiB
Markdown
88 lines
5.3 KiB
Markdown
# `[2026-09-17]` lv-bronte: corpus gated for real, adapter trained, SHIPPED with a FAILED voice axis on the record
|
||
|
||
**Status: SHIPPED 2026-09-17 01:24 as `lv-bronte` on `vllm-voices` (fv-ml1 GPU0 :8027), ckpt475 —
|
||
and it did NOT pass its voice gate.** Shipped because it is additive (one more named LoRA beside
|
||
`voices-base` and `lv-yarros`, reached only by requesting it), reversible (one compose line; hot-unload
|
||
measures 0.003 s), and clean on the SAFETY axis — 8-gram overlap identical to the never-saw-it control,
|
||
on a public-domain corpus. VRAM cost was nil: GPU0 96092 -> 96090 MiB. The caveat is written into the
|
||
compose file and into `/tank/aimodels/voice-adapters/lv-bronte-4b-v1/README.md` so it cannot be read
|
||
as a clean pass by anyone who finds the adapter without finding this note.
|
||
|
||
⚠ **Do NOT cite lv-bronte as evidence pair-SFT works for this author.** The voice axis is unresolved,
|
||
not passed.
|
||
|
||
## The gate result, in full
|
||
|
||
| axis | result | numbers |
|
||
|---|---|---|
|
||
| **A. VOICE** | ❌ **FAIL** (both candidates) | ckpt925 +0.210, ckpt475 +0.193 vs base — both **under** the 0.251 measured noise floor |
|
||
| **B. NOT COPIED** | ✅ PASS | ckpt475 **0.00 hit-rate, max 0 — identical to the never-saw-it control**; ckpt925 0.01, max 8. Positive control saturates at 160, so the detector is not blind |
|
||
| **C. NO DAMAGE** | ✅ PASS | ran-on +0.15 against a 0.400 floor |
|
||
|
||
```
|
||
same-author target (held-out Brontë vs itself) delta_cb 0.338 <- best achievable
|
||
ckpt925 0.531
|
||
ckpt475 0.548
|
||
base-unadapted 0.741
|
||
```
|
||
|
||
⭐ **THE EFFECT LOOKS REAL AND SUBSTANTIAL; THE INSTRUMENT CANNOT CERTIFY IT.** The reachable
|
||
span is 0.741 → 0.338 = 0.403, and the adapters closed **48–52% of everything achievable**.
|
||
Both beat base on *every individual seed*. This is an UNDERPOWERED result, not a null one —
|
||
and a "no effect" without its floor is unfalsifiable, so: **this method cannot resolve a voice
|
||
improvement smaller than ~0.251 delta_cb at 30 beats × 4 seeds on this corpus.**
|
||
|
||
⭐⭐ **THE CAUSE IS STRUCTURAL: Brontë's val split yields 81 pairs (44 in-band) against
|
||
Hemingway's 200**, because the corpus is 678k words against 994k. Maxing the fixture 30 → 44
|
||
beats would shrink the floor by only ~√1.47 ≈ 1.2× (to ~0.21, against a 0.21 gap) — still
|
||
marginal. **More SEEDS would not help either: the floor is a RANGE statistic, and ranges widen
|
||
with more samples.** There is no cheap fix.
|
||
|
||
## ⚠ A DEFECT IN THE v2 RULE ITSELF, worth fixing before the next author
|
||
|
||
The floor is defined as the **largest within-arm seed spread across ALL arms**. Measured here:
|
||
|
||
```
|
||
base-unadapted 0.772 0.813 0.751 0.772 spread 0.062
|
||
ckpt475 0.670 0.631 0.604 0.578 spread 0.092
|
||
ckpt925 0.776 0.584 0.525 0.620 spread 0.251 <- sets the floor, on ONE seed
|
||
```
|
||
|
||
So **adding a third, noisier arm raised the bar that failed the clean one.** Run as the
|
||
two-arm gate (base + ckpt475) the floor would have been 0.092 and +0.193 would have cleared
|
||
at 2.1×. This was NOT exploited — picking the floor that passes your preferred answer is the
|
||
threshold-chosen-after-seeing-the-numbers failure the pre-registration exists to prevent — but
|
||
the rule should say whether the floor is computed over the compared pair or over every arm
|
||
present. As written, a candidate's verdict depends on which *other* arms you happened to run.
|
||
|
||
**The outlier was diagnosed, not waved away.** Degeneracy probe (fraction of a generation made
|
||
of its most repeated 5-gram) is uniform across every seed and both arms, 0.0078–0.0102. Seed
|
||
1234 is not a collapsed generation; delta_cb genuinely has that variance for that arm.
|
||
|
||
## Which checkpoint, if it ships: **ckpt475**
|
||
|
||
The two are 0.017 apart on voice — far inside any floor, i.e. indistinguishable. On the axes
|
||
that DO resolve, ckpt475 wins both: memorisation identical to the control (ckpt925 has a
|
||
verbatim 8-gram hit), and **2.7× tighter seed-to-seed variance** (0.092 vs 0.251) with no
|
||
degeneracy to explain the difference — consistent with ckpt925 sitting nearer the overfit
|
||
boundary. Given a coin-flip on voice, take the one that provably did not memorise.
|
||
|
||
⭐ **THE RECIPE DID NOT TRANSFER.** Yarros and Hemingway both found their minimum inside
|
||
epoch two. Brontë's minima are step 475 (ep 1.00, 2.6107) and step 925 (ep 1.96, 2.6129) —
|
||
**0.0022 apart against a 0.0046 median neighbour jitter, i.e. indistinguishable**. Epoch 2
|
||
buys Brontë NOTHING over epoch 1. What IS robust is the epoch-3 collapse: +0.075, ~16× jitter.
|
||
Do not carry "two epochs on a three-epoch schedule" to a new author as settled.
|
||
|
||
## Artefacts
|
||
|
||
`gx10:~/lv-bronte/` (corpus-clean, corpus-renamed, entities-final.json, pairs/, beats-bronte-30.json),
|
||
`gx10:~/r49-runs/bronte-4b-pairs-3ep/` (57 checkpoints kept), `gx10:~/r49-runs/bronte-eval/`
|
||
(three arms × 120 generations, memorization.txt, voice_distance.txt, score.*.txt).
|
||
Commits `fc834a8` `533cc0c` `7964d07` `e9e8c40` `8bb7686`.
|
||
|
||
⚠ Two output labels in `voice_distance.py` are hardcoded Yarros strings — it prints
|
||
"reference: held-out Yarros" and a boilerplate "Base < Instruct" corroboration line. The
|
||
NUMBERS are Brontë's; those two labels are not. Not yet fixed.
|
||
|
||
Related: [[2026-09-16-lv-voices-line]], [[2026-09-16-lv-hemingway-corpus]], [[2026-09-16-voices-seat-lora]].
|