# SPDX-License-Identifier: Apache-2.0 # Copyright 2026 Kale """Single-sentence pool (15 Nov 2026): every sentence (3-28 words, first occurrence only) of the text pool's items (make_text.py output: its singles + every sentence of its passages, ban list already applied there). First used for m3's second batch (his kept takes were 86 % singles; passages failed pace / creak more often); make_jobs.py samples every later voice's singles from it. python3 -I make_singles.py work/jobs-v7.jsonl work/singles.jsonl "" [--voice m3] [--start 200000] Output rows: clip _, voice, text, description, kind "single", n_sent 1, source "pool sentence". Trimmed from the original: it also rebuilt each sentence's token ids through our earlier text front end's character table; ids now come from the release front end (frontend/) when the train step prepares features.""" import argparse import json import re from pathlib import Path ap = argparse.ArgumentParser() ap.add_argument("pool"); ap.add_argument("out"); ap.add_argument("description") ap.add_argument("--voice", default="m3"); ap.add_argument("--start", type=int, default=200000) A = ap.parse_args() jobs = [json.loads(l) for l in Path(A.pool).read_text().splitlines() if l.strip()] split = lambda t: [s for s in re.split(r"(?<=[.!?])\s+", t.strip()) if s] # noqa: E731 sents = [] seen = set() for j in jobs: for s in split(j["text"]): if s not in seen and 3 <= len(s.split()) <= 28: seen.add(s); sents.append(s) out = [] for n, s in enumerate(sents): c = f"{A.voice}_{A.start + n:06d}" out.append(dict(clip=c, voice=A.voice, text=s, description=A.description, mood=None, tagname=None, kind="single", n_sent=1, source="pool sentence")) Path(A.out).parent.mkdir(parents=True, exist_ok=True) Path(A.out).write_text("".join(json.dumps(x) + "\n" for x in out)) print(len(out), "single-sentence jobs")