обновлен процесс обработки данных

assembly_dataset/__main__.py

@@ -1,6 +1,7 @@
 import argparse
 import asyncio
 import json
+import signal
 from pathlib import Path
 import sys
 
@@ -8,6 +9,18 @@ from .config import load_config
 from .pipeline import generate, export, validate
 
 
+async def run_generation(config, *, show_progress, teacher_only=False):
+    loop = asyncio.get_running_loop()
+    task = asyncio.current_task()
+    previous = signal.getsignal(signal.SIGTERM)
+    loop.add_signal_handler(signal.SIGTERM, task.cancel)
+    try:
+        return await generate(config, show_progress=show_progress, teacher_only=teacher_only)
+    finally:
+        loop.remove_signal_handler(signal.SIGTERM)
+        signal.signal(signal.SIGTERM, previous)
+
+
 def main():
     parser = argparse.ArgumentParser(description='Датасет высокоуровневых планов сборки: full / step / branch')
     parser.add_argument('--config', type=Path, default=Path(__file__).resolve().parents[1] / 'config.yaml')
@@ -15,6 +28,8 @@ def main():
     parser.add_argument('--count', type=int, help='Точное суммарное число SFT-записей; 0 допустим')
     parser.add_argument('--output-dir', type=Path, help='Папка корпуса (относительно текущего каталога)')
     parser.add_argument('--dry-run', action='store_true', help='Показать настройки без записи и запросов')
+    parser.add_argument('--teacher-only', action='store_true', help='Сохранять только проверенные ответы LLM; останавливать генерацию при символической замене')
+    parser.add_argument('--allow-partial', action='store_true', help='Экспортировать накопленные записи, даже если цель ещё не достигнута')
     parser.add_argument('--no-progress', action='store_true', help='Отключить индикаторы прогресса')
     args = parser.parse_args()
     try:
@@ -32,11 +47,20 @@ def main():
             if config.teacher.enabled:
                 from dotenv import load_dotenv
                 load_dotenv(Path(__file__).resolve().parents[1] / '.env')
-            asyncio.run(generate(config, show_progress=not args.no_progress))
+            try:
+                asyncio.run(run_generation(config, show_progress=not args.no_progress, teacher_only=args.teacher_only))
+            except (Exception, KeyboardInterrupt, asyncio.CancelledError) as exc:
+                print(f'Генерация остановлена ({type(exc).__name__}). Экспорт накопленных данных…', file=sys.stderr)
+                try:
+                    saved = export(config, show_progress=not args.no_progress, allow_partial=True)
+                    print(f"Сохранено {saved['examples']} записей: {config.output_dir / 'sft'}", file=sys.stderr)
+                except Exception as export_error:
+                    print(f'Экспорт не выполнен ({type(export_error).__name__}); контрольные точки остаются в {config.output_dir / "episodes.sqlite3"}', file=sys.stderr)
+                return 130 if isinstance(exc, (KeyboardInterrupt, asyncio.CancelledError)) else 1
             result = export(config, show_progress=not args.no_progress)
             result = {k: result[k] for k in ('examples', 'scenarios', 'counts', 'teacher_failures')}
         elif args.command == 'export':
-            result = export(config, show_progress=not args.no_progress)
+            result = export(config, show_progress=not args.no_progress, allow_partial=args.allow_partial)
             result = {k: result[k] for k in ('examples', 'scenarios', 'counts', 'teacher_failures')}
         else:
             result = validate(config)

assembly_dataset/pipeline.py

@@ -53,12 +53,13 @@ def storage(config):
             db.close()
 
 
-async def generate(config: Config, agent=None, *, show_progress=False):
+async def generate(config: Config, agent=None, *, show_progress=False, teacher_only=False):
     with storage(config) as db:
         count = db.execute('SELECT COALESCE(SUM(rows), 0) FROM episodes').fetchone()[0]
         index = int(db.execute("SELECT value FROM meta WHERE key='next_index'").fetchone()[0])
+        resume_until = db.execute('SELECT COALESCE(MAX(seq), 0) FROM episodes').fetchone()[0]
         with ProgressDisplay('Генерация', config.target_examples, initial=count, enabled=show_progress) as progress:
-            if count >= config.target_examples:
+            if count >= config.target_examples and index >= resume_until:
                 return count
             if config.teacher.enabled and config.teacher.fraction > 0 and agent is None:
                 from .teacher import create_agent
@@ -66,28 +67,49 @@ async def generate(config: Config, agent=None, *, show_progress=False):
                 agent = await create_agent(config.teacher, config.prompts)
             batch_size = config.teacher.concurrency if agent is not None else 32
             duplicates = 0
-            while count < config.target_examples:
+            while count < config.target_examples or index < resume_until:
                 # Never construct 500k coroutines or keep the entire dataset in RAM.
                 size = min(batch_size, max(1, config.target_examples - count))
                 episodes = [make_episode(i, config) for i in range(index, index + size)]
+                async def save_episode(episode):
+                    nonlocal count, duplicates
+                    # A restarted batch can contain responses already committed before cancellation.
+                    if db.execute('SELECT 1 FROM episodes WHERE signature=?',
+                                  (episode['signature'],)).fetchone():
+                        duplicates += 1
+                        return
+                    if agent is not None:
+                        from .teacher import enrich
+                        episode, _ = await enrich(episode, config.teacher, agent, config.prompts)
+                    if teacher_only and (config.teacher.llm is None or
+                                         episode['source']['provider'] != config.teacher.llm.provider or
+                                         'teacher_error' in episode):
+                        raise ValueError('Режим --teacher-only: ответ модели не прошёл проверку; символическая замена не сохранена')
+                    rows = sum(1 for _ in training_rows(episode, config))
+                    # No await between receiving a validated response and committing it.
+                    with db:
+                        cursor = db.execute('INSERT OR IGNORE INTO episodes(seq,signature,payload,rows) VALUES (?,?,?,?)',
+                                            (episode['candidate_index'] + 1, episode['signature'], canonical(episode), rows))
+                    if cursor.rowcount:
+                        count += rows
+                        duplicates = 0
+                    else:
+                        duplicates += 1
+                    progress.update(count, status='ожидание LLM' if agent is not None else '')
+
                 if agent is not None:
-                    from .teacher import enrich
                     progress.update(status='ожидание LLM')
-                    episodes = [item[0] for item in await asyncio.gather(*(
-                        enrich(e, config.teacher, agent, config.prompts) for e in episodes))]
+                tasks = [asyncio.create_task(save_episode(episode)) for episode in episodes]
+                try:
+                    await asyncio.gather(*tasks)
+                finally:
+                    for task in tasks:
+                        if not task.done():
+                            task.cancel()
+                    await asyncio.gather(*tasks, return_exceptions=True)
+                # Advance only after the whole batch succeeds; holes are retried on resume.
+                index += size
                 with db:
-                    for episode in episodes:
-                        rows = sum(1 for _ in training_rows(episode, config))
-                        cursor = db.execute('INSERT OR IGNORE INTO episodes(signature,payload,rows) VALUES (?,?,?)',
-                                            (episode['signature'], canonical(episode), rows))
-                        if cursor.rowcount:
-                            count += rows
-                            duplicates = 0
-                        else:
-                            duplicates += 1
-                        index = episode['candidate_index'] + 1
-                        if count >= config.target_examples:
-                            break
                     db.execute("UPDATE meta SET value=? WHERE key='next_index'", (str(index),))
                 if duplicates > 10000:
                     raise ValueError('Пространство сценариев исчерпано: расширьте диапазон компонентов или domains')
@@ -140,10 +162,10 @@ def file_digest(path):
     return hasher.hexdigest()
 
 
-def export(config, *, show_progress=False):
+def export(config, *, show_progress=False, allow_partial=False):
     with storage(config) as db, ProgressDisplay('Экспорт', config.target_examples, enabled=show_progress) as progress:
         capacity = db.execute('SELECT COALESCE(SUM(rows), 0) FROM episodes').fetchone()[0]
-        if capacity < config.target_examples:
+        if capacity < config.target_examples and not allow_partial:
             raise ValueError('Недостаточно сценариев: сначала выполните generate')
         destination = config.output_dir / 'sft'
         backup = config.output_dir / '.sft-previous'
@@ -175,6 +197,7 @@ def export(config, *, show_progress=False):
             progress.update(status="контрольные суммы")
             writer.close()
             manifest = {'schema_version': VERSION, 'examples': total, 'scenarios': scenarios,
+                        'partial': total < config.target_examples,
                         'examples_per_variant': {'text': total, 'structured': total},
                         'counts': {f'{split}/{mode}': n for (split, mode), n in sorted(writer.counts.items())},
                         'domains': dict(domains), 'sources': dict(sources), 'teacher_failures': failures,

assembly_dataset/teacher.py

@@ -4,6 +4,7 @@ import re
 from typing import Literal
 
 from pydantic import Field, field_validator
+from pydantic_ai.exceptions import UnexpectedModelBehavior
 from llm.config import ConfigModel
 from .prompts import render
 from .language import validate_texts, action_text
@@ -63,6 +64,8 @@ async def enrich(episode, config, agent, prompts):
         except Exception as exc:
             # API exceptions may contain request headers. Persist the class only.
             error = {'scenario_id': episode['scenario_id'], 'error_type': type(exc).__name__, 'attempts': attempt + 1}
+            if not isinstance(exc, (ValueError, UnexpectedModelBehavior)) and attempt == config.retries:
+                raise RuntimeError(f'Запрос LLM не выполнен: {type(exc).__name__}. Генерация остановлена.') from None
             if isinstance(exc, ValueError):
                 prompt = render(prompts.teacher_retry, canonical({'instruction': episode['instruction'], 'scene': teacher_scene,
                                     'validation_error': str(exc)[:500]}))