Datasets:

idefinity
/

webfm-eval

Languages:

Chinese

Tags:

text-to-speech

evaluation

webfm

License:

Dataset card Files Files and versions

xet

Community

webfm-eval / kaggle /director_lora_data.mjs

zziang

Upload kaggle/director_lora_data.mjs with huggingface_hub

50e9802 verified about 13 hours ago

Raw History Blame Contribute Delete

19.6 kB

// Build and verify the director LoRA corpus against the LIVE browser protocol.
//
// The radio-script SFT set (kaggle/director_probe.py MODE=collect) teaches six
// lines of "开场：... / 叙述：...". The browser never asks for that. It asks for
// [memory] + [plan] + [continue]/[pause] + 0-3 spoken lines, and it parses the
// answer with director.js. Training on the radio format would teach the model
// to emit labels that cleanSpoken() then strips or throws away.
//
// So prompts come from the real Director class, and targets are accepted only
// if the real Director.parse() says they are clean. Same ruler on both sides.
//
// Usage (run locally, then push the outputs into the HF bundle):
//   node kaggle/director_lora_data.mjs prompts --n 4000 --out kaggle/out/director_prompts.jsonl
//   node kaggle/director_lora_data.mjs verify \
//        --prompts kaggle/out/director_prompts.jsonl \
//        --raw kaggle/out/director_raw.jsonl \
//        --out kaggle/out/director_sft.jsonl
//
// `raw` is one JSON object per line: {"id": "...", "raw": "<model output>"}.
// The generator of `raw` (base self-sampling or a larger teacher) lives on
// Kaggle; this script only decides what is good enough to train on.

import fs from 'node:fs';
import path from 'node:path';
import { fileURLToPath } from 'node:url';
import { Director } from '../director.js';
import { STORIES, CHAT } from './director_lora_corpus.mjs';

const HERE = path.dirname(fileURLToPath(import.meta.url));
const ROOT = path.resolve(HERE, '..');
const GENRES = JSON.parse(fs.readFileSync(path.join(ROOT, 'config/genres.json'), 'utf8')).genres;

const TOPICS = [
  '雨夜的末班公交', '一家很老的唱片店', '海边的灯塔', '凌晨的便利店',
  '搬家时翻出的旧照片', '楼下那棵一直没有开花的树', '深夜的收音机',
  '一场没有下完的雨', '旧公寓的电梯', '凌晨四点的菜市场', '山顶的观景台',
  '一个人的火锅店', '医院走廊的灯', '江边的长椅', '落了灰的钢琴',
  '雨天的出租车', '深夜的洗衣房', '开在二楼的旧书店', '废弃的火车站',
  '冬天的第一个霜降', '阳台上的那盆绿萝', '深夜的加油站', '清晨的面包房',
  '海边的小旅馆', '老街的修表摊', '天台上的风', '没有寄出的信',
  '一只总在窗边的猫', '雨停之后的马路', '深夜的游泳池', '断电的夜晚',
  '巷子尽头的路灯', '夏天最后一场雨', '老式闹钟的指针', '凌晨的公交站',
  '一条很长的隧道', '停电的电梯间', '小城的电影院', '合上的日记本',
  '厨房里烧开的水', '旧毛衣的口袋', '修好的那盏台灯', '雪停后的清晨',
];

const QUESTIONS = [
  '外面还在下雨吗', '你为什么总放这首歌', '现在几点了',
  '能说说海边吗', '你有没有想念过谁', '这封信后来寄出去了吗',
  '灯怎么忽然暗了', '你是谁', '今天有点累', '睡不着',
  '陪我说会儿话', '那个人后来回来了吗', '窗台上那盆花开了吗',
  '你还记得那家旧书店吗', '外面的车声怎么停了', '这首歌叫什么名字',
  '雨会下到天亮吗', '刚才那句话是什么意思', '你也听过钟声吗',
];

// These flip Director state in app.js (NEW_STORY_RE / CONTINUE_STORY_RE /
// STOP_RE / RESUME_RE), so a plain chat prompt would be off-distribution.
const STATE_FLIP = /讲(?:个|一个|一段)?故事|说(?:个|一个|一段)?故事|睡前故事|换个故事|新故事|继续讲|接着讲|不要停|晚安|别说了|回到剧本|继续读|接着读|停一下/;
const PLAIN_QUESTIONS = QUESTIONS.filter((q) => !STATE_FLIP.test(q));

const STORY_OPENERS = [
  '讲个故事吧', '说个故事', '来段睡前故事', '换个故事', '讲一个关于旧车站的故事',
];

const PLAN_SETS = [
  ['她在末班车站等一个不会来的人', '雨把时刻表泡得看不清', '一个卖花的老太太收摊',
   '她替老太太拎起铁桶', '末班车终于亮着灯拐过街角'],
  ['灯塔守夜人发现灯芯裂了', '他翻出三十年前的修理手册', '海面漂来一只空木箱',
   '箱子里是一叠没寄出的信', '天亮前灯重新亮起来'],
  ['旧书店在清仓', '男孩买到一本写满批注的书', '批注和他在同一页停住',
   '他按书里的地址找过去', '门牌还在，人已经搬走很久'],
  ['停电那晚整栋楼都醒了', '有人把蜡烛放在楼梯口', '邻居们第一次互相报名字',
   '电来的那一刻谁都没动', '蜡烛一直留到烧完'],
  ['修表摊的老匠人收了最后一个徒弟', '徒弟想换掉所有旧零件', '老匠人让他先听声音',
   '第一只被修好的表走慢了两分钟', '他们决定不调回来'],
];

const DOC_SNIPPET = '书页片段：《城与年》第七章\n'
  + '雨季过去的第三天，河滩上留下许多细小的贝壳。她沿着水线慢慢走，'
  + '鞋跟陷进湿沙里，每一步都发出很轻的声响。远处的摆渡人把船拴好，'
  + '抬头看了一眼云，又低下头去。';

const SPOKEN_HISTORY = [
  '灯亮起来的时候，站台只剩她一个。',
  '她把伞收好，听见远处有车进站。',
  '雨顺着棚檐滴下来，一滴一滴落在铁轨上。',
  '卖花的老太太把最后一把白菊抱进怀里。',
  '他翻过那页，纸角上还留着一小块水渍。',
  '走廊尽头的窗没关严，风把窗帘推得轻轻晃。',
];

function rng(seed) {
  let s = seed >>> 0;
  return () => {
    s = (s * 1664525 + 1013904223) >>> 0;
    return s / 4294967296;
  };
}

function pick(r, arr) { return arr[Math.floor(r() * arr.length) % arr.length]; }

function newDirector(genre) {
  const d = new Director({ genres: GENRES });
  d.setGenre(genre);
  return d;
}

function seedHistory(d, r, n = 3) {
  for (let i = 0; i < n; i++) {
    d.history.push({ role: i % 2 ? 'assistant' : 'user',
                     text: pick(r, i % 2 ? SPOKEN_HISTORY : PLAIN_QUESTIONS) });
  }
  d.memory = pick(r, [
    '她在末班车站等人，雨还没停。',
    '灯塔的灯芯裂了，守夜人正在翻旧手册。',
    '旧书店清仓，男孩买到一本写满批注的书。',
  ]);
}

function startStory(d, r, beat) {
  d.beginStory();
  d.program.plan = pick(r, PLAN_SETS);
  d.program.planAsked = true;
  d.program.beat = beat;
  d.memory = `故事进行到第 ${beat + 1} 拍：${d.program.plan[Math.min(beat, 4)]}`;
  seedHistory(d, r, 2 + Math.floor(r() * 3));
}

function storyFirst(d, r) {
  d.beginStory();
  return { userText: pick(r, STORY_OPENERS), autonomous: false };
}

// One build() per scenario. Each mutates a fresh Director into the state the
// app would be in, then returns the exact arguments Director.next() passes to
// messages().
const SCENES = [
  { name: 'chat_idle', genre: 'sleep', weight: 14, max: 280,
    build: (d, r) => { if (r() < 0.5) seedHistory(d, r); return { autonomous: true }; } },
  { name: 'chat_question', genre: 'sleep', weight: 14, max: 280,
    build: (d, r) => { if (r() < 0.6) seedHistory(d, r, 2);
      return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false }; } },
  { name: 'chat_followup', genre: 'sleep', weight: 8, max: 280,
    build: (d, r) => { seedHistory(d, r, 5); return { autonomous: true }; } },
  { name: 'chat_strict', genre: 'sleep', weight: 6, max: 240,
    build: (d, r) => { seedHistory(d, r, 2);
      return { autonomous: true, strict: true }; } },
  { name: 'story_plan', genre: 'story', weight: 10, max: 440,
    build: storyFirst },
  { name: 'story_beat', genre: 'story', weight: 16, max: 340,
    build: (d, r) => { startStory(d, r, 1 + Math.floor(r() * 20));
      return { autonomous: true }; } },
  { name: 'story_continue', genre: 'story', weight: 8, max: 340,
    build: (d, r) => { startStory(d, r, 2 + Math.floor(r() * 12));
      return { userText: '继续讲', autonomous: false }; } },
  { name: 'story_barge', genre: 'story', weight: 5, max: 340,
    build: (d, r) => { startStory(d, r, 3 + Math.floor(r() * 10));
      return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false }; } },
  { name: 'story_strict', genre: 'story', weight: 4, max: 240,
    build: (d, r) => { startStory(d, r, 4 + Math.floor(r() * 8));
      return { autonomous: true, strict: true }; } },
  { name: 'document_idle', genre: 'sleep', weight: 5, max: 280,
    build: (d, r) => { seedHistory(d, r, 2);
      return { autonomous: true, extraContext: DOC_SNIPPET }; } },
  { name: 'document_question', genre: 'sleep', weight: 5, max: 280,
    build: (d, r) => { seedHistory(d, r, 2);
      return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false,
               extraContext: DOC_SNIPPET }; } },
  { name: 'document_resume', genre: 'sleep', weight: 5, max: 280,
    build: (d, r) => { seedHistory(d, r, 3);
      return { userText: '回到剧本', autonomous: false,
               extraContext: `你在朗读《城与年》。\n${DOC_SNIPPET}` }; } },
];

function arg(name, fallback) {
  const i = process.argv.indexOf(`--${name}`);
  return i > 0 ? process.argv[i + 1] : fallback;
}

function readJsonl(file) {
  return fs.readFileSync(file, 'utf8').split('\n')
    .filter((l) => l.trim()).map((l) => JSON.parse(l));
}

function writeJsonl(file, rows) {
  fs.mkdirSync(path.dirname(file), { recursive: true });
  fs.writeFileSync(file, rows.map((r) => JSON.stringify(r, null, 0)).join('\n') + '\n');
}

function prompts() {
  const n = Number(arg('n', 4000));
  const seed = Number(arg('seed', 20261106));
  const out = arg('out', path.join(HERE, 'out', 'director_prompts.jsonl'));
  const r = rng(seed);
  const total = SCENES.reduce((s, x) => s + x.weight, 0);
  const rows = [];
  const counts = {};
  for (let i = 0; i < n; i++) {
    let t = r() * total, s = SCENES[0];
    for (const cand of SCENES) { t -= cand.weight; if (t <= 0) { s = cand; break; } }
    const d = newDirector(s.genre);
    const opts = s.build(d, r);
    const messages = d.messages(opts);
    counts[s.name] = (counts[s.name] || 0) + 1;
    rows.push({
      id: `d${String(i).padStart(5, '0')}`,
      kind: s.name,
      genre: s.genre,
      beat: d.program.beat,
      plan_wanted: d.wantsPlan(),
      max_new_tokens: s.max,
      messages,
    });
  }
  writeJsonl(out, rows);
  console.log(`wrote ${rows.length} prompts -> ${out}`);
  console.log(counts);
}

// Strict on purpose. cleanSpoken() in the app repairs dirty output at playback
// time; a training target must NOT need that repair, or the model learns to
// rely on it and the dirty-format rate goes up instead of down.
const DIRTY = /```|\|.*\||^\s*[-*#>]|你[：:]|用户[：:]|旁白[：:]|叙述者[：:]|作为\s*AI|<\|[^|>]*\|>|（[^）]*）|\([^)]*\)|(?:用户|对话|会话|节目)(?:已|结束)/m;

function judge(prompt, raw) {
  const d = newDirector(prompt.genre);
  // llm.worker.js decodes with skip_special_tokens:false, Kaggle samplers do
  // not. Drop chat markers at the edges; the trainer appends EOS itself.
  const text = String(raw || '').replace(/<\|[^|>]*\|>/g, ' ').trim();
  const p = d.parse(text, d.maxChars());
  const bad = [];
  const strict = prompt.kind.endsWith('_strict');
  if (!text) bad.push('empty');
  if (!p.lines.length) bad.push('no_lines');
  if (p.lines.length > 3) bad.push('too_many_lines');
  if (p.pause) bad.push('early_pause');
  if (!strict && !p.memory) bad.push('no_memory');
  else if (p.memory && !d.goodMemory(p.memory)) bad.push('meta_memory');
  if (!strict && !p.explicitContinue) bad.push('no_control');
  if (strict && /\[(?:memory|continue|pause|plan)\]/i.test(text)) bad.push('strict_has_tags');
  if (prompt.plan_wanted && p.plan.length < 5) bad.push('missing_plan');
  if (DIRTY.test(text)) bad.push('dirty');
  if (/[\u4e00-\u9fff]/.test(text) === false) bad.push('not_chinese');
  if (/[A-Za-z0-9]/.test(p.lines.join(''))) bad.push('latin_or_digit');
  const keys = p.lines.map((s) => s.replace(/[\s，,。！？!?、；;…：:]/g, ''));
  if (new Set(keys).size !== keys.length) bad.push('repeated_line');
  for (const line of p.lines) {
    if (line.length > d.maxChars() + 2) bad.push('long_line');
  }
  return { bad, parsed: p, text };
}

function verify() {
  const promptsFile = arg('prompts', path.join(HERE, 'out', 'director_prompts.jsonl'));
  const rawFile = arg('raw', path.join(HERE, 'out', 'director_raw.jsonl'));
  const out = arg('out', path.join(HERE, 'out', 'director_sft.jsonl'));
  const allPrompts = readJsonl(promptsFile);
  const byId = new Map(allPrompts.map((p) => [p.id, p]));
  const raws = readJsonl(rawFile);
  const kept = [];
  const seenResponse = new Set();
  const reasons = {};
  const dedup = process.argv.includes('--dedup');
  let missing = 0, dupes = 0;
  for (const row of raws) {
    const prompt = byId.get(row.id);
    if (!prompt) { missing++; continue; }
    const { bad, text } = judge(prompt, row.raw);
    if (bad.length) {
      for (const b of bad) reasons[b] = (reasons[b] || 0) + 1;
      continue;
    }
    if (dedup && seenResponse.has(text)) { dupes++; continue; }
    seenResponse.add(text);
    kept.push({ id: row.id, kind: prompt.kind, genre: prompt.genre,
                messages: prompt.messages,
                response: text });
  }
  let final = kept;
  const target = Number(arg('target', 0));
  if (target > 0) {
    const kindCount = {};
    for (const p of allPrompts) kindCount[p.kind] = (kindCount[p.kind] || 0) + 1;
    const groups = {};
    for (const row of kept) (groups[row.kind] ||= []).push(row);
    final = [];
    for (const [kind, rows] of Object.entries(groups)) {
      const quota = Math.max(1, Math.round(target * (kindCount[kind] || 0) / allPrompts.length));
      final.push(...rows.slice(0, quota));
    }
    const sr = rng(Number(arg('seed', 20261107)));
    for (let i = final.length - 1; i > 0; i--) {
      const j = Math.floor(sr() * (i + 1));
      [final[i], final[j]] = [final[j], final[i]];
    }
  }
  writeJsonl(out, final);
  const finalKinds = {};
  for (const row of final) finalKinds[row.kind] = (finalKinds[row.kind] || 0) + 1;
  console.log(`kept ${final.length} (from ${raws.length} raw, ${kept.length} clean, ${dupes} dupes dropped, ${missing} missing id)`);
  console.log('by kind', finalKinds);
  if (Object.keys(reasons).length) console.log('rejects', reasons);
}

function findStory(sys) {
  const m = sys.match(/计划：\n1\.\s*([^\n]+)/);
  if (!m) return null;
  return STORIES.find((s) => s.plan[0] === m[1].trim()) || null;
}

function pickVar(r, arr, k) { return arr[(Math.floor(r() * arr.length) + k) % arr.length]; }

// Free recombination at sentence level. Each sentence keeps the memory of the
// group it came from, so the [memory] line still matches the beat. For the
// question pool only group-opening sentences may lead the answer.
function pairsOf(pool, headed) {
  return {
    all: pool.flatMap((x) => x.lines.map((line, i) => ({
      line, memory: x.memory, head: headed && i === 0,
    }))),
    heads: pool.map((x) => ({ line: x.lines[0], memory: x.memory, head: true })),
  };
}

const CHAT_PAIRS = {
  idle: pairsOf(CHAT.idle, false),
  question: pairsOf(CHAT.question, true),
  document: pairsOf(CHAT.document, false),
  strict: pairsOf([...CHAT.idle, ...CHAT.question, ...CHAT.document,
                   ...CHAT.strict.map((lines) => ({ memory: '', lines }))], false),
};

function composeFree(r, pairSet, variant) {
  const pool = pairSet.heads.length ? pairSet.heads : pairSet.all;
  const i = Math.floor(r() * pool.length) + variant * 5;
  const a = pool[i % pool.length];
  const j = Math.floor(r() * pairSet.all.length) + variant * 11;
  const c = pairSet.all[j % pairSet.all.length];
  const lines = (variant % 5 === 4) ? [a.line]
    : [...new Set([a.line, c.line])];
  // Keep [memory] off the spoken line. Otherwise the model learns to write the
  // state twice: once as a tag, once aloud as the first sentence.
  const memPool = [...new Set(pairSet.all.map((x) => x.memory).filter(Boolean))];
  let memory = a.memory;
  for (let t = 0; t < 5 && memPool.length > 1; t++) {
    const cand = memPool[(Math.floor(r() * memPool.length) + t) % memPool.length];
    if (cand !== a.memory) { memory = cand; break; }
  }
  return { memory, lines };
}

function envelope(memory, lines) {
  return `[memory] ${memory}\n[continue]\n${lines.join('\n')}`;
}

// Rebuild a beat from a three-beat window. Sentences stay in narrative order,
// but which ones land in this answer changes, so 2k rows do not collapse into
// a couple hundred memorised strings.
function storyWindow(s, idx) {
  const out = [];
  for (let j = idx - 1; j <= idx + 1; j++) {
    out.push(s.beats[(j + s.beats.length) % s.beats.length]);
  }
  return out;
}

function composeWindow(r, items, variant) {
  const pool = items.flatMap((x) => x.lines);
  const count = 2 + (variant % 2);
  const start = (Math.floor(r() * 3) + variant * 2) % (pool.length - count + 1);
  const memory = items[(variant + 1) % items.length].memory;
  return envelope(memory, pool.slice(start, start + count));
}

function responseFor(p, r, variant) {
  const sys = p.messages[0].content;
  if (p.kind.endsWith('_strict')) {
    const c = composeFree(r, CHAT_PAIRS.strict, variant);
    return c.lines.join('\n');
  }
  if (p.kind === 'story_plan') {
    const s = pickVar(r, STORIES, variant);
    const o = pickVar(r, s.openings, variant);
    const pool = [...o.lines, ...storyWindow(s, variant % 3).flatMap((x) => x.lines)];
    const start = (Math.floor(r() * 4) + variant) % (pool.length - 1);
    const spoken = pool.slice(start, start + 2);
    return `[memory] ${o.memory}\n${s.plan.map((x) => `[plan] ${x}`).join('\n')}\n`
      + `[continue]\n${spoken.join('\n')}`;
  }
  if (p.kind.startsWith('story_')) {
    const s = findStory(sys) || pickVar(r, STORIES, variant);
    const m = sys.match(/第 (\d+) 拍/);
    const beat = m ? Number(m[1]) : 2;
    const idx = (Math.max(1, beat) - 1 + variant) % s.beats.length;
    return composeWindow(r, storyWindow(s, idx), variant);
  }
  const pairSet = p.kind.startsWith('document_') ? CHAT_PAIRS.document
    : p.kind === 'chat_question' ? CHAT_PAIRS.question : CHAT_PAIRS.idle;
  const c = composeFree(r, pairSet, variant);
  return envelope(c.memory, c.lines);
}

function assemble() {
  const promptsFile = arg('prompts', path.join(HERE, 'out', 'director_prompts.jsonl'));
  const out = arg('out', path.join(HERE, 'out', 'director_raw.jsonl'));
  const seed = Number(arg('seed', 20261107));
  const rows = readJsonl(promptsFile);
  const r = rng(seed);
  const seen = new Set();
  const raws = [];
  const rejects = {};
  let accepted = 0;
  for (const p of rows) {
    let raw = '';
    for (let k = 0; k < 10; k++) {
      raw = responseFor(p, r, k);
      if (!seen.has(raw)) break;
    }
    seen.add(raw);
    const { bad } = judge(p, raw);
    if (bad.length) for (const x of bad) rejects[x] = (rejects[x] || 0) + 1;
    else accepted++;
    raws.push({ id: p.id, raw });
  }
  writeJsonl(out, raws);
  console.log(`assembled ${raws.length} responses -> ${out}`);
  console.log(`unique ${seen.size} (${(seen.size / raws.length * 100).toFixed(1)}%)`);
  console.log(`verify preview: ${accepted}/${raws.length} (${(accepted / raws.length * 100).toFixed(1)}%) would pass`);
  if (Object.keys(rejects).length) console.log('rejects', rejects);
}

const cmd = process.argv[2];
if (cmd === 'prompts') prompts();
else if (cmd === 'assemble') assemble();
else if (cmd === 'verify') verify();
else {
  console.log('usage: director_lora_data.mjs prompts|assemble|verify [--n --seed --out --prompts --raw]');
  process.exit(1);
}

	// Build and verify the director LoRA corpus against the LIVE browser protocol.
	//
	// The radio-script SFT set (kaggle/director_probe.py MODE=collect) teaches six
	// lines of "开场：... / 叙述：...". The browser never asks for that. It asks for
	// [memory] + [plan] + [continue]/[pause] + 0-3 spoken lines, and it parses the
	// answer with director.js. Training on the radio format would teach the model
	// to emit labels that cleanSpoken() then strips or throws away.
	//
	// So prompts come from the real Director class, and targets are accepted only
	// if the real Director.parse() says they are clean. Same ruler on both sides.
	//
	// Usage (run locally, then push the outputs into the HF bundle):
	// node kaggle/director_lora_data.mjs prompts --n 4000 --out kaggle/out/director_prompts.jsonl
	// node kaggle/director_lora_data.mjs verify \
	// --prompts kaggle/out/director_prompts.jsonl \
	// --raw kaggle/out/director_raw.jsonl \
	// --out kaggle/out/director_sft.jsonl
	//
	// `raw` is one JSON object per line: {"id": "...", "raw": "<model output>"}.
	// The generator of `raw` (base self-sampling or a larger teacher) lives on
	// Kaggle; this script only decides what is good enough to train on.

	import fs from 'node:fs';
	import path from 'node:path';
	import { fileURLToPath } from 'node:url';
	import { Director } from '../director.js';
	import { STORIES, CHAT } from './director_lora_corpus.mjs';

	const HERE = path.dirname(fileURLToPath(import.meta.url));
	const ROOT = path.resolve(HERE, '..');
	const GENRES = JSON.parse(fs.readFileSync(path.join(ROOT, 'config/genres.json'), 'utf8')).genres;

	const TOPICS = [
	'雨夜的末班公交', '一家很老的唱片店', '海边的灯塔', '凌晨的便利店',
	'搬家时翻出的旧照片', '楼下那棵一直没有开花的树', '深夜的收音机',
	'一场没有下完的雨', '旧公寓的电梯', '凌晨四点的菜市场', '山顶的观景台',
	'一个人的火锅店', '医院走廊的灯', '江边的长椅', '落了灰的钢琴',
	'雨天的出租车', '深夜的洗衣房', '开在二楼的旧书店', '废弃的火车站',
	'冬天的第一个霜降', '阳台上的那盆绿萝', '深夜的加油站', '清晨的面包房',
	'海边的小旅馆', '老街的修表摊', '天台上的风', '没有寄出的信',
	'一只总在窗边的猫', '雨停之后的马路', '深夜的游泳池', '断电的夜晚',
	'巷子尽头的路灯', '夏天最后一场雨', '老式闹钟的指针', '凌晨的公交站',
	'一条很长的隧道', '停电的电梯间', '小城的电影院', '合上的日记本',
	'厨房里烧开的水', '旧毛衣的口袋', '修好的那盏台灯', '雪停后的清晨',
	];

	const QUESTIONS = [
	'外面还在下雨吗', '你为什么总放这首歌', '现在几点了',
	'能说说海边吗', '你有没有想念过谁', '这封信后来寄出去了吗',
	'灯怎么忽然暗了', '你是谁', '今天有点累', '睡不着',
	'陪我说会儿话', '那个人后来回来了吗', '窗台上那盆花开了吗',
	'你还记得那家旧书店吗', '外面的车声怎么停了', '这首歌叫什么名字',
	'雨会下到天亮吗', '刚才那句话是什么意思', '你也听过钟声吗',
	];

	// These flip Director state in app.js (NEW_STORY_RE / CONTINUE_STORY_RE /
	// STOP_RE / RESUME_RE), so a plain chat prompt would be off-distribution.
	const STATE_FLIP = /讲(?:个\|一个\|一段)?故事\|说(?:个\|一个\|一段)?故事\|睡前故事\|换个故事\|新故事\|继续讲\|接着讲\|不要停\|晚安\|别说了\|回到剧本\|继续读\|接着读\|停一下/;
	const PLAIN_QUESTIONS = QUESTIONS.filter((q) => !STATE_FLIP.test(q));

	const STORY_OPENERS = [
	'讲个故事吧', '说个故事', '来段睡前故事', '换个故事', '讲一个关于旧车站的故事',
	];

	const PLAN_SETS = [
	['她在末班车站等一个不会来的人', '雨把时刻表泡得看不清', '一个卖花的老太太收摊',
	'她替老太太拎起铁桶', '末班车终于亮着灯拐过街角'],
	['灯塔守夜人发现灯芯裂了', '他翻出三十年前的修理手册', '海面漂来一只空木箱',
	'箱子里是一叠没寄出的信', '天亮前灯重新亮起来'],
	['旧书店在清仓', '男孩买到一本写满批注的书', '批注和他在同一页停住',
	'他按书里的地址找过去', '门牌还在，人已经搬走很久'],
	['停电那晚整栋楼都醒了', '有人把蜡烛放在楼梯口', '邻居们第一次互相报名字',
	'电来的那一刻谁都没动', '蜡烛一直留到烧完'],
	['修表摊的老匠人收了最后一个徒弟', '徒弟想换掉所有旧零件', '老匠人让他先听声音',
	'第一只被修好的表走慢了两分钟', '他们决定不调回来'],
	];

	const DOC_SNIPPET = '书页片段：《城与年》第七章\n'
	+ '雨季过去的第三天，河滩上留下许多细小的贝壳。她沿着水线慢慢走，'
	+ '鞋跟陷进湿沙里，每一步都发出很轻的声响。远处的摆渡人把船拴好，'
	+ '抬头看了一眼云，又低下头去。';

	const SPOKEN_HISTORY = [
	'灯亮起来的时候，站台只剩她一个。',
	'她把伞收好，听见远处有车进站。',
	'雨顺着棚檐滴下来，一滴一滴落在铁轨上。',
	'卖花的老太太把最后一把白菊抱进怀里。',
	'他翻过那页，纸角上还留着一小块水渍。',
	'走廊尽头的窗没关严，风把窗帘推得轻轻晃。',
	];

	function rng(seed) {
	let s = seed >>> 0;
	return () => {
	s = (s * 1664525 + 1013904223) >>> 0;
	return s / 4294967296;
	};
	}

	function pick(r, arr) { return arr[Math.floor(r() * arr.length) % arr.length]; }

	function newDirector(genre) {
	const d = new Director({ genres: GENRES });
	d.setGenre(genre);
	return d;
	}

	function seedHistory(d, r, n = 3) {
	for (let i = 0; i < n; i++) {
	d.history.push({ role: i % 2 ? 'assistant' : 'user',
	text: pick(r, i % 2 ? SPOKEN_HISTORY : PLAIN_QUESTIONS) });
	}
	d.memory = pick(r, [
	'她在末班车站等人，雨还没停。',
	'灯塔的灯芯裂了，守夜人正在翻旧手册。',
	'旧书店清仓，男孩买到一本写满批注的书。',
	]);
	}

	function startStory(d, r, beat) {
	d.beginStory();
	d.program.plan = pick(r, PLAN_SETS);
	d.program.planAsked = true;
	d.program.beat = beat;
	d.memory = `故事进行到第 ${beat + 1} 拍：${d.program.plan[Math.min(beat, 4)]}`;
	seedHistory(d, r, 2 + Math.floor(r() * 3));
	}

	function storyFirst(d, r) {
	d.beginStory();
	return { userText: pick(r, STORY_OPENERS), autonomous: false };
	}

	// One build() per scenario. Each mutates a fresh Director into the state the
	// app would be in, then returns the exact arguments Director.next() passes to
	// messages().
	const SCENES = [
	{ name: 'chat_idle', genre: 'sleep', weight: 14, max: 280,
	build: (d, r) => { if (r() < 0.5) seedHistory(d, r); return { autonomous: true }; } },
	{ name: 'chat_question', genre: 'sleep', weight: 14, max: 280,
	build: (d, r) => { if (r() < 0.6) seedHistory(d, r, 2);
	return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false }; } },
	{ name: 'chat_followup', genre: 'sleep', weight: 8, max: 280,
	build: (d, r) => { seedHistory(d, r, 5); return { autonomous: true }; } },
	{ name: 'chat_strict', genre: 'sleep', weight: 6, max: 240,
	build: (d, r) => { seedHistory(d, r, 2);
	return { autonomous: true, strict: true }; } },
	{ name: 'story_plan', genre: 'story', weight: 10, max: 440,
	build: storyFirst },
	{ name: 'story_beat', genre: 'story', weight: 16, max: 340,
	build: (d, r) => { startStory(d, r, 1 + Math.floor(r() * 20));
	return { autonomous: true }; } },
	{ name: 'story_continue', genre: 'story', weight: 8, max: 340,
	build: (d, r) => { startStory(d, r, 2 + Math.floor(r() * 12));
	return { userText: '继续讲', autonomous: false }; } },
	{ name: 'story_barge', genre: 'story', weight: 5, max: 340,
	build: (d, r) => { startStory(d, r, 3 + Math.floor(r() * 10));
	return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false }; } },
	{ name: 'story_strict', genre: 'story', weight: 4, max: 240,
	build: (d, r) => { startStory(d, r, 4 + Math.floor(r() * 8));
	return { autonomous: true, strict: true }; } },
	{ name: 'document_idle', genre: 'sleep', weight: 5, max: 280,
	build: (d, r) => { seedHistory(d, r, 2);
	return { autonomous: true, extraContext: DOC_SNIPPET }; } },
	{ name: 'document_question', genre: 'sleep', weight: 5, max: 280,
	build: (d, r) => { seedHistory(d, r, 2);
	return { userText: pick(r, PLAIN_QUESTIONS), autonomous: false,
	extraContext: DOC_SNIPPET }; } },
	{ name: 'document_resume', genre: 'sleep', weight: 5, max: 280,
	build: (d, r) => { seedHistory(d, r, 3);
	return { userText: '回到剧本', autonomous: false,
	extraContext: `你在朗读《城与年》。\n${DOC_SNIPPET}` }; } },
	];

	function arg(name, fallback) {
	const i = process.argv.indexOf(`--${name}`);
	return i > 0 ? process.argv[i + 1] : fallback;
	}

	function readJsonl(file) {
	return fs.readFileSync(file, 'utf8').split('\n')
	.filter((l) => l.trim()).map((l) => JSON.parse(l));
	}

	function writeJsonl(file, rows) {
	fs.mkdirSync(path.dirname(file), { recursive: true });
	fs.writeFileSync(file, rows.map((r) => JSON.stringify(r, null, 0)).join('\n') + '\n');
	}

	function prompts() {
	const n = Number(arg('n', 4000));
	const seed = Number(arg('seed', 20261106));
	const out = arg('out', path.join(HERE, 'out', 'director_prompts.jsonl'));
	const r = rng(seed);
	const total = SCENES.reduce((s, x) => s + x.weight, 0);
	const rows = [];
	const counts = {};
	for (let i = 0; i < n; i++) {
	let t = r() * total, s = SCENES[0];
	for (const cand of SCENES) { t -= cand.weight; if (t <= 0) { s = cand; break; } }
	const d = newDirector(s.genre);
	const opts = s.build(d, r);
	const messages = d.messages(opts);
	counts[s.name] = (counts[s.name] \|\| 0) + 1;
	rows.push({
	id: `d${String(i).padStart(5, '0')}`,
	kind: s.name,
	genre: s.genre,
	beat: d.program.beat,
	plan_wanted: d.wantsPlan(),
	max_new_tokens: s.max,
	messages,
	});
	}
	writeJsonl(out, rows);
	console.log(`wrote ${rows.length} prompts -> ${out}`);
	console.log(counts);
	}

	// Strict on purpose. cleanSpoken() in the app repairs dirty output at playback
	// time; a training target must NOT need that repair, or the model learns to
	// rely on it and the dirty-format rate goes up instead of down.
	const DIRTY = /```\|\\|.\\|\|^\s[-#>]\|你[：:]\|用户[：:]\|旁白[：:]\|叙述者[：:]\|作为\sAI\|<\\|[^\|>]\\|>\|（[^）]）\|\([^)]*\)\|(?:用户\|对话\|会话\|节目)(?:已\|结束)/m;

	function judge(prompt, raw) {
	const d = newDirector(prompt.genre);
	// llm.worker.js decodes with skip_special_tokens:false, Kaggle samplers do
	// not. Drop chat markers at the edges; the trainer appends EOS itself.
	const text = String(raw \|\| '').replace(/<\\|[^\|>]*\\|>/g, ' ').trim();
	const p = d.parse(text, d.maxChars());
	const bad = [];
	const strict = prompt.kind.endsWith('_strict');
	if (!text) bad.push('empty');
	if (!p.lines.length) bad.push('no_lines');
	if (p.lines.length > 3) bad.push('too_many_lines');
	if (p.pause) bad.push('early_pause');
	if (!strict && !p.memory) bad.push('no_memory');
	else if (p.memory && !d.goodMemory(p.memory)) bad.push('meta_memory');
	if (!strict && !p.explicitContinue) bad.push('no_control');
	if (strict && /\[(?:memory\|continue\|pause\|plan)\]/i.test(text)) bad.push('strict_has_tags');
	if (prompt.plan_wanted && p.plan.length < 5) bad.push('missing_plan');
	if (DIRTY.test(text)) bad.push('dirty');
	if (/[\u4e00-\u9fff]/.test(text) === false) bad.push('not_chinese');
	if (/[A-Za-z0-9]/.test(p.lines.join(''))) bad.push('latin_or_digit');
	const keys = p.lines.map((s) => s.replace(/[\s，,。！？!?、；;…：:]/g, ''));
	if (new Set(keys).size !== keys.length) bad.push('repeated_line');
	for (const line of p.lines) {
	if (line.length > d.maxChars() + 2) bad.push('long_line');
	}
	return { bad, parsed: p, text };
	}

	function verify() {
	const promptsFile = arg('prompts', path.join(HERE, 'out', 'director_prompts.jsonl'));
	const rawFile = arg('raw', path.join(HERE, 'out', 'director_raw.jsonl'));
	const out = arg('out', path.join(HERE, 'out', 'director_sft.jsonl'));
	const allPrompts = readJsonl(promptsFile);
	const byId = new Map(allPrompts.map((p) => [p.id, p]));
	const raws = readJsonl(rawFile);
	const kept = [];
	const seenResponse = new Set();
	const reasons = {};
	const dedup = process.argv.includes('--dedup');
	let missing = 0, dupes = 0;
	for (const row of raws) {
	const prompt = byId.get(row.id);
	if (!prompt) { missing++; continue; }
	const { bad, text } = judge(prompt, row.raw);
	if (bad.length) {
	for (const b of bad) reasons[b] = (reasons[b] \|\| 0) + 1;
	continue;
	}
	if (dedup && seenResponse.has(text)) { dupes++; continue; }
	seenResponse.add(text);
	kept.push({ id: row.id, kind: prompt.kind, genre: prompt.genre,
	messages: prompt.messages,
	response: text });
	}
	let final = kept;
	const target = Number(arg('target', 0));
	if (target > 0) {
	const kindCount = {};
	for (const p of allPrompts) kindCount[p.kind] = (kindCount[p.kind] \|\| 0) + 1;
	const groups = {};
	for (const row of kept) (groups[row.kind] \|\|= []).push(row);
	final = [];
	for (const [kind, rows] of Object.entries(groups)) {
	const quota = Math.max(1, Math.round(target * (kindCount[kind] \|\| 0) / allPrompts.length));
	final.push(...rows.slice(0, quota));
	}
	const sr = rng(Number(arg('seed', 20261107)));
	for (let i = final.length - 1; i > 0; i--) {
	const j = Math.floor(sr() * (i + 1));
	[final[i], final[j]] = [final[j], final[i]];
	}
	}
	writeJsonl(out, final);
	const finalKinds = {};
	for (const row of final) finalKinds[row.kind] = (finalKinds[row.kind] \|\| 0) + 1;
	console.log(`kept ${final.length} (from ${raws.length} raw, ${kept.length} clean, ${dupes} dupes dropped, ${missing} missing id)`);
	console.log('by kind', finalKinds);
	if (Object.keys(reasons).length) console.log('rejects', reasons);
	}

	function findStory(sys) {
	const m = sys.match(/计划：\n1\.\s*([^\n]+)/);
	if (!m) return null;
	return STORIES.find((s) => s.plan[0] === m[1].trim()) \|\| null;
	}

	function pickVar(r, arr, k) { return arr[(Math.floor(r() * arr.length) + k) % arr.length]; }

	// Free recombination at sentence level. Each sentence keeps the memory of the
	// group it came from, so the [memory] line still matches the beat. For the
	// question pool only group-opening sentences may lead the answer.
	function pairsOf(pool, headed) {
	return {
	all: pool.flatMap((x) => x.lines.map((line, i) => ({
	line, memory: x.memory, head: headed && i === 0,
	}))),
	heads: pool.map((x) => ({ line: x.lines[0], memory: x.memory, head: true })),
	};
	}

	const CHAT_PAIRS = {
	idle: pairsOf(CHAT.idle, false),
	question: pairsOf(CHAT.question, true),
	document: pairsOf(CHAT.document, false),
	strict: pairsOf([...CHAT.idle, ...CHAT.question, ...CHAT.document,
	...CHAT.strict.map((lines) => ({ memory: '', lines }))], false),
	};

	function composeFree(r, pairSet, variant) {
	const pool = pairSet.heads.length ? pairSet.heads : pairSet.all;
	const i = Math.floor(r() * pool.length) + variant * 5;
	const a = pool[i % pool.length];
	const j = Math.floor(r() * pairSet.all.length) + variant * 11;
	const c = pairSet.all[j % pairSet.all.length];
	const lines = (variant % 5 === 4) ? [a.line]
	: [...new Set([a.line, c.line])];
	// Keep [memory] off the spoken line. Otherwise the model learns to write the
	// state twice: once as a tag, once aloud as the first sentence.
	const memPool = [...new Set(pairSet.all.map((x) => x.memory).filter(Boolean))];
	let memory = a.memory;
	for (let t = 0; t < 5 && memPool.length > 1; t++) {
	const cand = memPool[(Math.floor(r() * memPool.length) + t) % memPool.length];
	if (cand !== a.memory) { memory = cand; break; }
	}
	return { memory, lines };
	}

	function envelope(memory, lines) {
	return `[memory] ${memory}\n[continue]\n${lines.join('\n')}`;
	}

	// Rebuild a beat from a three-beat window. Sentences stay in narrative order,
	// but which ones land in this answer changes, so 2k rows do not collapse into
	// a couple hundred memorised strings.
	function storyWindow(s, idx) {
	const out = [];
	for (let j = idx - 1; j <= idx + 1; j++) {
	out.push(s.beats[(j + s.beats.length) % s.beats.length]);
	}
	return out;
	}

	function composeWindow(r, items, variant) {
	const pool = items.flatMap((x) => x.lines);
	const count = 2 + (variant % 2);
	const start = (Math.floor(r() * 3) + variant * 2) % (pool.length - count + 1);
	const memory = items[(variant + 1) % items.length].memory;
	return envelope(memory, pool.slice(start, start + count));
	}

	function responseFor(p, r, variant) {
	const sys = p.messages[0].content;
	if (p.kind.endsWith('_strict')) {
	const c = composeFree(r, CHAT_PAIRS.strict, variant);
	return c.lines.join('\n');
	}
	if (p.kind === 'story_plan') {
	const s = pickVar(r, STORIES, variant);
	const o = pickVar(r, s.openings, variant);
	const pool = [...o.lines, ...storyWindow(s, variant % 3).flatMap((x) => x.lines)];
	const start = (Math.floor(r() * 4) + variant) % (pool.length - 1);
	const spoken = pool.slice(start, start + 2);
	return `[memory] ${o.memory}\n${s.plan.map((x) => `[plan] ${x}`).join('\n')}\n`
	+ `[continue]\n${spoken.join('\n')}`;
	}
	if (p.kind.startsWith('story_')) {
	const s = findStory(sys) \|\| pickVar(r, STORIES, variant);
	const m = sys.match(/第 (\d+) 拍/);
	const beat = m ? Number(m[1]) : 2;
	const idx = (Math.max(1, beat) - 1 + variant) % s.beats.length;
	return composeWindow(r, storyWindow(s, idx), variant);
	}
	const pairSet = p.kind.startsWith('document_') ? CHAT_PAIRS.document
	: p.kind === 'chat_question' ? CHAT_PAIRS.question : CHAT_PAIRS.idle;
	const c = composeFree(r, pairSet, variant);
	return envelope(c.memory, c.lines);
	}

	function assemble() {
	const promptsFile = arg('prompts', path.join(HERE, 'out', 'director_prompts.jsonl'));
	const out = arg('out', path.join(HERE, 'out', 'director_raw.jsonl'));
	const seed = Number(arg('seed', 20261107));
	const rows = readJsonl(promptsFile);
	const r = rng(seed);
	const seen = new Set();
	const raws = [];
	const rejects = {};
	let accepted = 0;
	for (const p of rows) {
	let raw = '';
	for (let k = 0; k < 10; k++) {
	raw = responseFor(p, r, k);
	if (!seen.has(raw)) break;
	}
	seen.add(raw);
	const { bad } = judge(p, raw);
	if (bad.length) for (const x of bad) rejects[x] = (rejects[x] \|\| 0) + 1;
	else accepted++;
	raws.push({ id: p.id, raw });
	}
	writeJsonl(out, raws);
	console.log(`assembled ${raws.length} responses -> ${out}`);
	console.log(`unique ${seen.size} (${(seen.size / raws.length * 100).toFixed(1)}%)`);
	console.log(`verify preview: ${accepted}/${raws.length} (${(accepted / raws.length * 100).toFixed(1)}%) would pass`);
	if (Object.keys(rejects).length) console.log('rejects', rejects);
	}

	const cmd = process.argv[2];
	if (cmd === 'prompts') prompts();
	else if (cmd === 'assemble') assemble();
	else if (cmd === 'verify') verify();
	else {
	console.log('usage: director_lora_data.mjs prompts\|assemble\|verify [--n --seed --out --prompts --raw]');
	process.exit(1);
	}