From e9e8c40b83295aa37576f9f17be2e3b8f23ab500 Mon Sep 17 00:00:00 2001 From: Vuong Hoang Date: Wed, 16 Sep 2026 21:18:44 -0700 Subject: [PATCH] eval harness: sample the beat fixture from held-out val, and bind the eval prompt to the trained one MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Two harness defects that would each make a voice number uninterpretable. build_beat_fixture.py — the fixture is now SAMPLED from the val split rather than hand-written. The original BabyYarros fixture was five hand-written beats about a stray dog and a kitten: wrong genre, so 'He licked her clean' came back as explicit sex from a romantasy adapter, and n=5 had a noise floor of 0.800 that manufactured a +0.45 result which collapsed to +0.08 at n=120. Sampling from val makes it in-genre and held out by construction, spread across works so a naive head(30) is not one novel. Refuses outright if the pairs carry any split but val, because a fixture drawn from training data makes every downstream number a memorisation measurement wearing a voice label. gen_beats_chat_yarros.py --system-from — the SYS constant in this harness is Yarros's. Driving a Bronte or Hemingway adapter with it measures the arm under a system prompt it was never trained on and confounds the carrier change with a prompt change. Rather than duplicate the register table and rely on whoever runs it to pick the matching one, read the prompt out of the pair build's own provenance, which is the artefact that records what the adapter actually saw. --- scripts/r49-corpus/build_beat_fixture.py | 93 +++++++++++++++++++++ scripts/r49-corpus/gen_beats_chat_yarros.py | 19 +++++ 2 files changed, 112 insertions(+) create mode 100644 scripts/r49-corpus/build_beat_fixture.py diff --git a/scripts/r49-corpus/build_beat_fixture.py b/scripts/r49-corpus/build_beat_fixture.py new file mode 100644 index 0000000..0704eec --- /dev/null +++ b/scripts/r49-corpus/build_beat_fixture.py @@ -0,0 +1,93 @@ +"""Build the N-beat evaluation fixture from HELD-OUT val pairs. + +⚠ THE FIXTURE MUST BE IN-GENRE, AND THAT IS NOT A STYLE PREFERENCE. The first +BabyYarros fixture was five hand-written beats about a stray dog and a kitten. +Two things went wrong and both were invisible until measured: + + * WRONG GENRE. `He licked her clean, and would not be driven off` came back as + explicit sex from an adapter trained on romantasy. A beat that never occurs in + the training distribution measures the carrier's priors, not the voice. + * n=5 HAS NO RESOLUTION. That fixture's noise floor was 0.800 and it manufactured + a +0.45 in-band "result" which collapsed to +0.08 — inside a 0.233 floor — at + n=120. One sample moves a rate by 0.2 when there are five. + +So the fixture is SAMPLED from the val split the adapter never trained on: in-genre +by construction, held out by construction, and large enough that the per-arm sample +(N beats x S seeds) can resolve something. Nothing here is hand-written. + +Emits the `[{"id","beat"}]` shape gen_beats_chat_*.py reads, plus a sidecar keeping +each beat's true held-out response so a memorisation check has its reference text. +""" +from __future__ import annotations +import argparse, json, random +from pathlib import Path + + +def main() -> int: + ap = argparse.ArgumentParser() + ap.add_argument("--pairs", required=True, help="held-out val pairs jsonl") + ap.add_argument("--out", required=True) + ap.add_argument("--sidecar", default=None, + help="write {id, beat, response, work, chapter} here for memorisation checks") + ap.add_argument("-n", type=int, default=30) + ap.add_argument("--seed", type=int, default=4919) + ap.add_argument("--min-words", type=int, default=90, + help="skip val passages outside the product band; a 40-word reference " + "cannot tell an in-band arm from a truncated one") + ap.add_argument("--max-words", type=int, default=150) + a = ap.parse_args() + + rows = [json.loads(l) for l in Path(a.pairs).read_text(encoding="utf-8").splitlines() if l.strip()] + if not rows: + print(f"== {a.pairs} is empty"); return 1 + + split = {r.get("split") for r in rows} + if split - {"val"}: + # Sampling the fixture from anything the adapter trained on makes every + # downstream number a memorisation measurement wearing a voice label. + print(f"== REFUSING: pairs carry split(s) {sorted(split)}; the fixture must be val-only") + return 1 + + band = [r for r in rows if a.min_words <= r.get("words", 0) <= a.max_words] + if len(band) < a.n: + print(f"== only {len(band)} val pairs inside {a.min_words}-{a.max_words} words, need {a.n}") + return 1 + + # Spread across works rather than taking a block: the val split is emitted in + # corpus order, so a naive head(30) would be one novel and one register. + by_work: dict[str, list] = {} + for r in band: + by_work.setdefault(r["work"], []).append(r) + rng = random.Random(a.seed) + for v in by_work.values(): + rng.shuffle(v) + + picked, works = [], sorted(by_work) + i = 0 + while len(picked) < a.n: + w = works[i % len(works)] + if by_work[w]: + picked.append(by_work[w].pop()) + elif not any(by_work.values()): + break + i += 1 + + fixture = [{"id": f"b{n+1}", "beat": r["beat"]} for n, r in enumerate(picked)] + Path(a.out).write_text(json.dumps(fixture, ensure_ascii=False, indent=1), encoding="utf-8") + + if a.sidecar: + side = [{"id": f"b{n+1}", "beat": r["beat"], "response": r["response"], + "work": r["work"], "chapter": r.get("chapter"), "words": r.get("words")} + for n, r in enumerate(picked)] + Path(a.sidecar).write_text(json.dumps(side, ensure_ascii=False, indent=1), encoding="utf-8") + + dist: dict[str, int] = {} + for r in picked: + dist[r["work"]] = dist.get(r["work"], 0) + 1 + print(f" {len(fixture)} beats from {len(band)} in-band val pairs · per work: {dist}") + print(f" wrote {a.out}" + (f" + {a.sidecar}" if a.sidecar else "")) + return 0 + + +if __name__ == "__main__": + raise SystemExit(main()) diff --git a/scripts/r49-corpus/gen_beats_chat_yarros.py b/scripts/r49-corpus/gen_beats_chat_yarros.py index 8787191..9b31da3 100644 --- a/scripts/r49-corpus/gen_beats_chat_yarros.py +++ b/scripts/r49-corpus/gen_beats_chat_yarros.py @@ -44,8 +44,27 @@ ap.add_argument("--top-p", type=float, default=0.95) # arm stays byte-reproducible; the flag exists so the casing can be MEASURED as its own # variable instead of being confounded with the adapter it is being used to judge. ap.add_argument("--user-prefix", default="BEAT: ") +# ⭐ TRAIN/EVAL PROMPT PARITY, MADE STRUCTURAL RATHER THAN REMEMBERED. +# The hardcoded SYS above is Yarros's. Driving a Brontë or Hemingway adapter with it +# would confound the carrier change with a PROMPT change -- the arm would be measured +# under a system prompt it was never trained on, and the resulting delta would be +# uninterpretable. Rather than duplicate the register table here and rely on whoever +# runs this to pick the matching one, read the prompt straight out of the pair build's +# own provenance, which is the artefact that records what the adapter actually saw. +ap.add_argument("--system-from", default=None, + help="pairs provenance json; its `system_prompt` replaces SYS. Use this for " + "any corpus but Yarros — it guarantees the eval drives the adapter under " + "the prompt it was trained on.") a = ap.parse_args() +if a.system_from: + _prov = json.loads(Path(a.system_from).read_text()) + _sp = _prov.get("system_prompt") + if not _sp: + raise SystemExit(f"== {a.system_from} carries no `system_prompt` -- refusing to guess") + SYS = _sp + print(f" system prompt from {a.system_from}: {SYS[:80]}...") + beats = json.loads(Path(a.beats).read_text()) tok = AutoTokenizer.from_pretrained(a.base) if tok.chat_template is None: