обновлен формат генерации датасета

tests/test_assembly_dataset.py

@@ -0,0 +1,114 @@
+"""Behavioral tests for exact sizes, restart safety and logical plan validation."""
+import asyncio
+import copy
+import json
+import tempfile
+import unittest
+from pathlib import Path
+
+
+class AssemblyTests(unittest.TestCase):
+    def test_preinstalled_parts_do_not_depend_on_unfinished_assembly(self):
+        from assembly_dataset.world import make_episode, RECIPES
+        from assembly_dataset.config import Config
+        for index in range(100):
+            episode = make_episode(index, Config())
+            scene = episode['scene']
+            last = len(RECIPES[scene['domain']][1]) - 1
+            for component in scene['components']:
+                if f"{component['id']}:stage:{last}" in scene['initial_state']:
+                    self.assertEqual(component['dependencies'], [])
+
+    def test_exact_counts_resume_and_split_grouping(self):
+        from assembly_dataset.config import Config
+        from assembly_dataset.pipeline import generate, export
+        for count in (0, 1, 2, 7, 100):
+            with self.subTest(count=count), tempfile.TemporaryDirectory() as tmp:
+                config = Config(output_dir=Path(tmp), target_examples=count, shard_size=11)
+                asyncio.run(generate(config))
+                export(config)
+                rows = [json.loads(line) for p in Path(tmp).glob('sft/*/*/[0-9][0-9][0-9][0-9][0-9][0-9].jsonl')
+                        for line in p.read_text().splitlines()]
+                self.assertEqual(len(rows), count)
+                self.assertEqual(len({r['record_id'] for r in rows}), count)
+                groups = {}
+                for r in rows:
+                    self.assertEqual(groups.setdefault(r['scenario_id'], r['split']), r['split'])
+                asyncio.run(generate(config))
+                export(config)
+                self.assertEqual(sum(len(p.read_text().splitlines()) for p in Path(tmp).glob('sft/*/*/[0-9][0-9][0-9][0-9][0-9][0-9].jsonl')), count)
+                config.target_examples = count + 4
+                asyncio.run(generate(config))
+                export(config)
+                self.assertEqual(sum(len(p.read_text().splitlines()) for p in Path(tmp).glob('sft/*/*/[0-9][0-9][0-9][0-9][0-9][0-9].jsonl')), count + 4)
+
+    def test_rejects_changed_generation_settings_on_resume(self):
+        from assembly_dataset.config import Config
+        from assembly_dataset.pipeline import generate
+        with tempfile.TemporaryDirectory() as tmp:
+            config = Config(output_dir=Path(tmp), target_examples=4)
+            asyncio.run(generate(config))
+            config.seed += 1
+            with self.assertRaisesRegex(ValueError, 'config|конфиг'):
+                asyncio.run(generate(config))
+
+    def test_plans_and_branches_are_executable(self):
+        from assembly_dataset.world import make_episode, validate_plan, validate_branch
+        from assembly_dataset.config import Config
+        for index in range(100):
+            episode = make_episode(index, Config())
+            validate_plan(episode, episode['plan'])
+            validate_branch(episode, episode['branch'])
+            if episode['plan']['actions']:
+                invalid = copy.deepcopy(episode['plan'])
+                invalid['actions'][0]['object'] = 'does_not_exist'
+                with self.assertRaises(ValueError):
+                    validate_plan(episode, invalid)
+                invalid_branch = copy.deepcopy(episode['branch'])
+                invalid_branch['nodes'][0]['on_failure'] = 'missing_node'
+                with self.assertRaises(ValueError):
+                    validate_branch(episode, invalid_branch)
+
+    def test_missing_goal_and_cycles_are_rejected(self):
+        from assembly_dataset.world import make_episode, validate_plan, validate_branch
+        from assembly_dataset.config import Config
+        episode = next(e for i in range(100) if (e := make_episode(i, Config()))['plan']['status'] == 'ready')
+        broken = copy.deepcopy(episode['plan'])
+        broken['actions'] = []
+        with self.assertRaises(ValueError):
+            validate_plan(episode, broken)
+        branch = copy.deepcopy(episode['branch'])
+        branch['nodes'][0]['on_success'] = branch['nodes'][0]['id']
+        with self.assertRaises(ValueError):
+            validate_branch(episode, branch)
+
+    def test_failure_branch_requires_fresh_scene_before_another_action(self):
+        from assembly_dataset.world import make_episode, validate_branch
+        from assembly_dataset.config import Config
+        episode = make_episode(0, Config(blocked_fraction=0))
+        branch = copy.deepcopy(episode['branch'])
+        retry = copy.deepcopy(branch['nodes'][0])
+        retry['id'] = 'retry'
+        branch['nodes'][0]['on_failure'] = 'retry'
+        branch['nodes'].append(retry)
+        with self.assertRaises(ValueError):
+            validate_branch(episode, branch)
+
+    def test_reproducibility_and_structural_diversity(self):
+        from assembly_dataset.world import make_episode
+        from assembly_dataset.config import Config
+        config = Config()
+        signatures = {make_episode(i, config)['signature'] for i in range(1000)}
+        self.assertGreater(len(signatures), 990)
+        self.assertEqual(make_episode(12, config), make_episode(12, config))
+
+    def test_config_rejects_invalid_counts_modes_and_splits(self):
+        from assembly_dataset.config import Config
+        for kwargs in ({'target_examples': -1}, {'target_examples': 1.5}, {'modes': []},
+                       {'modes': ['full', 'full']}, {'split': [0.8, 0.5, 0.1]}):
+            with self.subTest(kwargs=kwargs), self.assertRaises(ValueError):
+                Config(**kwargs)
+
+
+if __name__ == '__main__':
+    unittest.main()