| |
| """Meaningful language integrity checks, independently of format/physics checks.""" |
| from pathlib import Path |
| import json, re, importlib.util |
| P=Path(__file__).resolve().parent |
| spec=importlib.util.spec_from_file_location('normalize',P/'normalize_language.py');m=importlib.util.module_from_spec(spec);spec.loader.exec_module(m) |
| checks=[] |
| def check(name,condition): |
| checks.append({'check':name,'passed':bool(condition)}) |
| if not condition:raise AssertionError(name) |
| M=json.loads((P/'exact_instruction_map.json').read_text()) |
| for s,r in M.items(): |
| n=m.normalize_instruction(s) |
| check('strict_exact:'+s[:60],not m.audit_instruction(n) and n==r['normalized_instruction']) |
| check('idempotent:'+s[:60],m.normalize_instruction(n)==n) |
| check('orange_and_blue_are_distinct',m.normalize_instruction('Pick up gatorade_m.')=='Pick up orange sports drink.' and m.normalize_instruction('Pick up gatorade_t.')=='Pick up blue sports drink.') |
| for unknown in ['Pick up Gatorade.','Pick up /World/Assets/CoffeeTable/UnknownItem.','Pick up unknown_asset_41.','请做一个未提供翻译的任务。']: |
| try:m.normalize_instruction(unknown) |
| except m.LanguageNormalizationError:check('unresolved_rejected:'+unknown,True) |
| else:check('unresolved_rejected:'+unknown,False) |
| check('known_prim_wrapper_resolved',m.normalize_instruction('Pick up /World/RoomEpisode/CakeYellow/Asset/Object.')=='Pick up yellow pastry.') |
| nested={'people':['No coffee for the children.',{'restriction':'Do not serve snacks containing peanuts.','age':[3,6]}],'goals':('Pick up gatorade_m.','Leave gatorade_t in place.'),'optional':None,'count':2} |
| out=m.normalize_context(nested) |
| check('nested_context_types_and_all_leaves_preserved',list(out)==list(nested) and out['people']==nested['people'] and out['optional'] is None and out['count']==2 and isinstance(out['goals'],tuple) and out['goals']==('Pick up orange sports drink.','Leave blue sports drink in place.')) |
| long_open=json.loads((P/'long_open_language.json').read_text())['episodes'] |
| for r in long_open: |
| n=r['normalized_instruction'] |
| if r['task_kind']=='open' and r.get('context_provenance')!='proposed_after_execution_not_original_user_intent':check('recorded_context_materialized:'+r['episode_id'],n.startswith('Context:\n') and '\n\nGoal:\n' in n) |
| if r.get('context_provenance')=='proposed_after_execution_not_original_user_intent':check('posthoc_context_provenance_only:'+r['episode_id'],r['training_context'] is None and r['source_delivered_posthoc_context']==r['source_context'] and bool(r['normalized_delivered_posthoc_context']) and not n.startswith('Context:')) |
| if r['source_id']=='tidy_coffee_table_v3': |
| src=r['source_instruction'].split('桌上有',1)[1].split('，把它们',1)[0] |
| expected=re.findall(r'\d+',src) |
| actual=re.findall(r'\d+',n.split('It has ',1)[1].split('.',1)[0]) |
| check('quantities_preserved:'+r['episode_id'],expected==actual) |
| check('goals_and_protected_items_preserved:'+r['episode_id'],all(s in n for s in ['general waste bin','recycling bin','storage basket','drawer back','TV cabinet','unopened']) and 'drawer closed' not in n) |
| alltext='\n'.join(r['normalized_instruction'] for r in long_open) |
| check('guest_preferences_and_negation_preserved',all(s in alltext for s in ['ages 3 and 6','must not be served coffee or cola','cannot drink alcohol','do not serve any snacks containing peanuts','must be put out of view'])) |
| E=json.loads((P/'episode_instruction_map.json').read_text()) |
| for key,r in E.items(): |
| if r['source_id']!='long_pack6_choose4':continue |
| n=m.normalize_instruction(r['source_instruction'],episode_id=key) |
| check('bag_archive_executed_object_list:'+key,all(o in n for o in r['executed_semantic_objects']) and len(r['executed_semantic_objects'])==4) |
| check('bag_request_authority_truthful:'+key,':requested_objects' not in ' '.join(r['evidence']) and ((r['preexecution_goal_authority_verified'] is True and r.get('original_input_long_job') and r.get('requested_objects')==r['executed_objects']) or (r['preexecution_goal_authority_verified'] is False and 'requested_objects' not in r))) |
| check('ten_bag_episode_specific_maps',sum(r['source_id']=='long_pack6_choose4' for r in E.values())==10) |
| try:m.normalize_instruction(next(iter(E.values()))['source_instruction']) |
| except m.LanguageNormalizationError:check('generic_bag_instruction_cannot_silently_drop_object_list',True) |
| else:check('generic_bag_instruction_cannot_silently_drop_object_list',False) |
| meal_path=P/'meal_diagnostic_language.json' |
| if meal_path.exists(): |
| meal=json.loads(meal_path.read_text())['episodes'] |
| check('ten_meal_current_language_records',len(meal)==10) |
| for r in meal: |
| check('meal_no_invented_context_or_oracle:'+r['episode_id'],r['source_context'] is None and r['training_context'] is None and r['current_instruction_origin']=='user_adopted_delivery_entry_scope' and r['historical_source_success_flag'] is False and r['current_task_scope_status']=='passed' and r['release_role']=='train_candidate' and not any(x in r['normalized_instruction'].lower() for x in ['cutlery','knife','fork','failed','successful']) and m.compose_training_instruction(r['source_instruction'],episode_id=r['episode_id'])==r['normalized_instruction']) |
| bench20_path=P/'bench20_ready5_language.json' |
| if bench20_path.exists(): |
| rows=json.loads(bench20_path.read_text())['episodes'] |
| check('five_current_l05_structured_requests',len(rows)==5) |
| for r in rows: |
| check('l05_preexecution_goal_without_route_or_oracle:'+r['episode_id'],r['preexecution_goal_authority_verified'] is True and r['original_episode_instruction_missing'] is True and r['is_exact_owner_natural_language_utterance'] is False and r['training_context'] is None and r['normalized_instruction']==r['source_instruction'] and r['normalized_instruction'].startswith('Find the ') and not any(s in r['normalized_instruction'] for s in ['sofa','kitchen','coffee table','not found','successful','failed'])) |
| guest_probe=P/'CURRENT_REQUEST_GUEST10_AND_PACK_ROOTS.json' |
| if guest_probe.exists(): |
| rows=json.loads(guest_probe.read_text())['guest'] |
| check('guest_current_recipe_ten',len(rows)==10) |
| for r in rows: |
| check('guest_current_recipe_and_delivered_text_match:'+r['episode_id'],all(r.get(k) is True for k in ['recipe_matches_delivered_snapshot','context_matches_selected','instruction_matches_selected','situation_matches_selected'])) |
| report={'schema':'robotagentbench.semantic_language_validation.v1','checks':len(checks),'passed':sum(c['passed'] for c in checks),'all_passed':all(c['passed'] for c in checks),'details':checks,'limitations':['Lexical and preservation checks do not approve physical quality or owner acceptance.','Episode-specific bag maps require the source episode_id; a missing identifier is rejected.','Post-hoc Serve Tidy contexts retain source context_provenance; language does not turn them into preregistered scenarios.']} |
| (P/'LANGUAGE_VALIDATION.json').write_text(json.dumps(report,ensure_ascii=False,indent=2)+'\n') |
| print(json.dumps({k:v for k,v in report.items() if k!='details'},ensure_ascii=False,indent=2)) |
|
|
