EgoBench / analysis_scripts /analyze_error_reasons.py
emodiary's picture
Initial complete EgoBench release
3c4b9d3 verified
Raw History Blame Contribute Delete
42.3 kB
#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ All-Scenario Error Reason Analysis Script - Based on eval_result data Supported scenarios: retail, restaurant (including multi-restaurant variant 6), kitchen, warehouse, household Error reasons are determined for each sample in the following order: 1. Syntax format error 2. Multimodal recognition error (whether query tool calls match key/value) 3. Hallucination issue (incorrect user_id parameter) 4. Tool call logic reasoning error (matches < total_gt_calls, difference >= 2) 5. Memory/common sense error (matches < total_gt_calls, difference == 1) 6. Over-operation error (matches == total_gt_calls but result_based failed) Correct sample: both tool-based evaluation and result-based evaluation are correct """ import json import os import sys from collections import defaultdict from typing import Dict, List, Optional, Tuple sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..')) from tools.retail.retail_db import RetailDB from tools.retail.retail_init import ( retail_init_data1, retail_init_data2, retail_init_data3, retail_init_data4, retail_init_data5, retail_init_data6, retail_init_data7, retail_init_data8, retail_init_data9, retail_init_data10 ) from tools.restaurant.restaurant_db import RestaurantDB from tools.restaurant.restaurant_init import restaurant_init_data, restaurant_init_data5 from tools.restaurant.restaurant6_db import Restaurant6DB from tools.restaurant.restaurant6_init import restaurant6_init_data from tools.kitchen.kitchen_db import KitchenDB from tools.kitchen.kitchen_init import kitchen_init_data from tools.warehouse.warehouse_db import WarehouseDB from tools.warehouse import warehouse_init from tools.household.household_db import HouseholdDB from tools.household import household_init MODELS = [ 'qwen3-vl-235b', 'Qwen3.5-397B-A17B', 'gemini-3.1-pro-preview', 'kimi-k2.6', 'mimo-v2.5-omni', 'qwen3.6-plus', 'glm-5v-turbo', 'doubao-seed-2-0-pro-260215' ] # ===================== Configurable I/O Roots ===================== # Defaults preserve the original behaviour (reading results/ + eval_result/, # writing error_analysis/). Override via CLI (see main()) to analyse a # different run, e.g. the GPT-user run: # --results_root GPT_user_results --eval_root GPT_user_eval_result # --output_root error_analysis_gpt_user RESULTS_ROOT = 'results' EVAL_ROOT = 'eval_result' OUTPUT_ROOT = 'error_analysis' ACTIVE_MODELS = list(MODELS) INCLUDE_PARTIAL = False ERROR_TYPES = { 'syntax_error': 'Syntax format error in tool calls', 'multimodal_error': 'Multimodal recognition error', 'hallucination_error': 'Hallucination issue (incorrect user_id parameter)', 'logic_error': 'Tool call logic reasoning error', 'memory_error': 'Memory or common sense error', 'over_operation_error': 'Over-operation error', 'correct': 'Correct' } MODES = ['easy', 'hard', 'static'] SCENARIO_CONFIG = { 'retail': { 'scenario_numbers': list(range(1, 11)), 'init_data_map': { 1: retail_init_data1, 2: retail_init_data2, 3: retail_init_data3, 4: retail_init_data4, 5: retail_init_data5, 6: retail_init_data6, 7: retail_init_data7, 8: retail_init_data8, 9: retail_init_data9, 10: retail_init_data10 }, 'db_class': RetailDB, 'query_tools': { 'get_price', 'get_tax_rate', 'get_category', 'get_discount', 'get_nutrition', 'find_products_by_nutritional_characteristic', 'find_products_by_taste', 'find_products_by_country_of_origin', 'find_products_by_price_range', 'list_discounted_products' }, }, 'restaurant': { 'scenario_numbers': list(range(1, 7)), 'init_data_map': { 1: restaurant_init_data, 2: restaurant_init_data, 3: restaurant_init_data, 4: restaurant_init_data, 5: restaurant_init_data5, 6: restaurant6_init_data, }, 'db_class': RestaurantDB, 'db_class_map': {6: Restaurant6DB}, 'query_tools': { 'get_dish_price', 'get_dish_discount', 'get_dish_nutrition', 'get_dish_allergens', 'get_tax_rate', 'get_dish_taste_profile', 'get_user_order_summary', 'get_set_meal_details', 'find_dishes_by_category', 'find_dishes_by_nutritional_tag', 'find_dishes_by_taste', 'find_set_meals_containing_dish', 'list_all_discounted_dishes', 'filter_dishes_by_price_range', 'compute_total_payment', 'compute_total_tax', 'compute_total_nutrition' }, }, 'kitchen': { 'scenario_numbers': list(range(1, 5)), 'init_data_map': { 1: kitchen_init_data, 2: kitchen_init_data, 3: kitchen_init_data, 4: kitchen_init_data }, 'db_class': KitchenDB, 'query_tools': { 'get_current_menu', 'get_cooking_steps', 'get_recipe_allergens', 'get_recipe_taste', 'get_recipe_ingredients', 'get_ingredient_shelf_life', 'get_ingredient_location', 'get_ingredient_nutrition', 'get_ingredient_quantity', 'get_current_shopping_list', 'get_recipe_nutritional_characteristics', 'find_recipes_by_allergen', 'find_recipes_by_taste', 'find_recipes_by_ingredient', 'find_ingredients_by_expiry_date', 'find_ingredients_by_location', 'find_recipes_by_nutritional_characteristics', 'get_ingredients_by_category', 'find_ingredient_category', 'get_all_recipe_names', 'get_all_ingredient_names', 'tally_total_nutritional_characteristics', 'tally_total_tastes', 'compute_total_nutritions' }, }, 'warehouse': { 'scenario_numbers': list(range(1, 26)), 'init_data_map': { number: getattr(warehouse_init, f'warehouse_init_data{number}') for number in range(1, 26) }, 'db_class': WarehouseDB, 'query_tools': { 'get_all_equipment', 'get_equipment_category', 'find_equipment_by_category', 'get_equipment_brand', 'find_equipment_by_brand', 'get_equipment_country_of_origin', 'find_equipment_by_country_of_origin', 'get_equipment_purchase_date', 'find_equipment_by_purchase_date_range', 'get_equipment_ideal_storage_location', 'find_equipment_by_ideal_storage_location', 'get_equipment_usage_history', 'get_equipment_maintenance_and_usage_instructions', 'get_equipment_maintenance_history', 'get_equipment_action_list', }, }, 'household': { 'scenario_numbers': list(range(1, 19)), 'init_data_map': { number: getattr(household_init, f'household_init_data{number}') for number in range(1, 19) }, 'db_class': HouseholdDB, 'query_tools': { 'get_all_item_names', 'get_item_category', 'find_items_by_category', 'get_item_ideal_storage_location', 'find_items_by_ideal_storage_location', 'get_item_purchase_date', 'find_items_by_purchase_date_range', 'get_item_purchase_price', 'find_items_by_purchase_price_range', 'get_item_usage_history', 'find_items_used_during_date_range', 'get_item_maintenance_history', 'get_item_maintenance_and_usage_instructions', 'get_household_action_list', }, }, } ALL_SCENARIO_TYPES = list(SCENARIO_CONFIG.keys()) # Ground-truth annotations and tool schemas do not always use the same field # name for the visually grounded entity. In particular, household annotations # store the target under ``name``, while household query tools accept # ``item_name``. Treat these as aliases so a correctly grounded query is not # mislabeled as a multimodal recognition error. MULTIMODAL_PARAMETER_ALIASES = { 'household': { 'name': ('name', 'item_name'), }, } def load_init_db(scenario_type: str, scenario_number: int): """Initialize database based on scenario type and number""" config = SCENARIO_CONFIG[scenario_type] db_class = config.get('db_class_map', {}).get( scenario_number, config['db_class'] ) db = db_class() init_data = config['init_data_map'].get(scenario_number) if init_data: db.init_from_json(init_data) return db def load_ground_truth(scenario_type: str, scenario_number: int) -> List[Dict]: """Load ground truth file""" gt_path = os.path.join( os.path.dirname(__file__), '..', 'scenarios', 'final', f'{scenario_type}{scenario_number}.json' ) gt_path = os.path.normpath(gt_path) if not os.path.exists(gt_path): return [] with open(gt_path, 'r', encoding='utf-8') as f: return json.load(f) def load_interaction_log(model_name: str, scenario_type: str, scenario_number: int, mode: str) -> Optional[List[Dict]]: """Load interaction log""" log_path = os.path.join( os.path.dirname(__file__), '..', RESULTS_ROOT, model_name, f'{scenario_type}{scenario_number}_{mode}.json' ) log_path = os.path.normpath(log_path) if not os.path.exists(log_path): return None with open(log_path, 'r', encoding='utf-8') as f: return json.load(f) def load_eval_result(model_name: str, scenario_type: str, scenario_number: int, mode: str) -> Optional[Dict]: """Load evaluation result""" eval_path = os.path.join( os.path.dirname(__file__), '..', EVAL_ROOT, model_name, f'{scenario_type}{scenario_number}_{mode}_eval.json' ) eval_path = os.path.normpath(eval_path) if not os.path.exists(eval_path): return None with open(eval_path, 'r', encoding='utf-8') as f: return json.load(f) def is_complete_eval_result(eval_result: Dict) -> bool: """Return whether an evaluated result file covered every ground-truth task.""" if "is_complete" in eval_result: return bool(eval_result["is_complete"]) total = eval_result.get("total_scenarios", 0) valid = eval_result.get("valid_scenarios", 0) return total > 0 and valid == total and not eval_result.get("invalid_scenarios") def is_usable_eval_result(eval_result: Dict) -> bool: """Return whether an eval file may contribute to the selected analysis.""" if is_complete_eval_result(eval_result): return True return INCLUDE_PARTIAL and eval_result.get("valid_scenarios", 0) > 0 def extract_all_tool_calls(tool_calls: List[Dict]) -> List[Dict]: """Extract all tool calls (flatten)""" all_calls = [] for entry in tool_calls: if not isinstance(entry, dict): continue if "calls" in entry and isinstance(entry["calls"], list): for call in entry["calls"]: if isinstance(call, dict): all_calls.append({ "tool_name": call.get("tool_name") or call.get("name"), "parameters": call.get("parameters", {}) }) elif "call" in entry: call = entry.get("call", {}) if isinstance(call, dict): all_calls.append({ "tool_name": call.get("tool_name") or call.get("name"), "parameters": call.get("parameters", {}) }) return all_calls def check_syntax_error(tool_calls: List[Dict]) -> Tuple[bool, str]: """ 1. Check if tool call syntax format is incorrect Returns: (has_syntax_error, error_description) """ if not tool_calls: return True, "No tool calls" for i, call_entry in enumerate(tool_calls): if not isinstance(call_entry, dict): return True, f"Tool call entry {i + 1} has incorrect format" if "calls" in call_entry: calls = call_entry.get("calls", []) if not isinstance(calls, list): return True, f"Tool {i + 1} 'calls' field is not a list" for j, call in enumerate(calls): if not isinstance(call, dict): return True, f"Tool {i + 1} call {j + 1} has incorrect format" if "tool_name" not in call and "name" not in call: return True, f"Tool {i + 1} call {j + 1} missing tool_name" if "parameters" not in call: return True, f"Tool {i + 1} call {j + 1} missing parameters" elif "call" in call_entry: call = call_entry.get("call", {}) if not isinstance(call, dict): return True, f"Tool {i + 1} 'call' field has incorrect format" if "tool_name" not in call and "name" not in call: return True, f"Tool {i + 1} missing tool_name" if "parameters" not in call: return True, f"Tool {i + 1} missing parameters" else: return True, f"Tool call entry {i + 1} has incorrect format" return False, "" def check_multimodal_error( gt_entry: Dict, interaction_calls: List[Dict], query_tools: set, scenario_type: str = None, scenario_number: int = None ) -> Tuple[bool, str]: """ 2. Check multimodal recognition error By comparing whether all query tool calls match the multimodal information in the ground truth file, i.e., whether the tested agent's tool call parameters contain the key and value from the scenario file. There must be tested tool calls, and if the value list has multiple values, tool calls containing all these parameters must be present. For Restaurant 6, also check whether the selected restaurant is correct. Returns: (has_multimodal_error, error_description) """ # Restaurant 6 is the multi-restaurant selection scenario. if scenario_type == "restaurant" and scenario_number == 6: expected_restaurant = "Mediterranean Greek Restaurant" # Find the last tool call with restaurant_name parameter last_restaurant_choice = None for call in reversed(interaction_calls): params = call.get("parameters", {}) if "restaurant_name" in params and isinstance(params["restaurant_name"], str): last_restaurant_choice = params["restaurant_name"] break if last_restaurant_choice is not None and last_restaurant_choice != expected_restaurant: return True, f"Incorrect restaurant selection: expected '{expected_restaurant}', got '{last_restaurant_choice}'" key = gt_entry.get("key") value = gt_entry.get("value", []) if not key or not value: return False, "No multimodal recognition requirement" if not interaction_calls: return False, "No tool calls" parameter_keys = MULTIMODAL_PARAMETER_ALIASES.get( scenario_type, {} ).get(key, (key,)) matched_values = set() for call in interaction_calls: tool_name = call.get("tool_name", "") if tool_name not in query_tools: continue params = call.get("parameters", {}) parameter_key = next( (candidate for candidate in parameter_keys if candidate in params), None, ) if parameter_key is None: continue param_value = params[parameter_key] param_values = param_value if isinstance(param_value, list) else [param_value] for actual_value in param_values: for expected_value in value: if isinstance(actual_value, str) and isinstance(expected_value, str): if actual_value.lower().strip() == expected_value.lower().strip(): matched_values.add(expected_value) elif expected_value.lower() in actual_value.lower(): matched_values.add(expected_value) elif fuzzy_match_name(actual_value, expected_value): matched_values.add(expected_value) elif actual_value == expected_value: matched_values.add(expected_value) unmatched_values = set(value) - matched_values if unmatched_values: return True, f"Failed to recognize {key}={list(unmatched_values)}" return False, "" def fuzzy_match_name(param_value: str, expected_value: str) -> bool: """Fuzzy match name""" param_lower = param_value.lower().strip() expected_lower = expected_value.lower().strip() if param_lower in expected_lower or expected_lower in param_lower: return True param_words = set(param_lower.split()) expected_words = set(expected_lower.split()) return bool(param_words & expected_words) def check_hallucination_error(gt_entry: Dict, interaction_calls: List[Dict]) -> Tuple[bool, str]: """ 3. Check hallucination error (incorrect user_id parameter) Returns: (has_hallucination_error, error_description) """ correct_user_ids = { call.get("parameters", {}).get("user_id") for call in gt_entry.get("ground_truth", []) if call.get("parameters", {}).get("user_id") is not None } if not correct_user_ids: return False, "No user_id requirement" wrong_user_ids = [] for call in interaction_calls: params = call.get("parameters", {}) if "user_id" in params and params["user_id"] not in correct_user_ids: wrong_user_ids.append(params["user_id"]) if wrong_user_ids: expected = sorted(correct_user_ids) actual = sorted(set(wrong_user_ids)) if len(expected) == 1: return True, f"user_id should be '{expected[0]}', but found {actual}" return True, f"user_id should be one of {expected}, but found {actual}" return False, "" def analyze_scenario( scenario_type: str, gt_entry: Dict, interaction_entry: Dict, eval_detail: Dict, scenario_number: int ) -> Dict: """ Analyze error reasons for a single scenario First determine if successful (tool_based_success AND result_based_success), if successful, classify as correct; if failed, determine error reasons in order: 1. Syntax error -> 2. Multimodal error -> 3. Hallucination error -> 4. Logic error -> 5. Memory error -> 6. Over-operation """ result = { "scenario_type": scenario_type, "scenario_number": scenario_number, "scenario_id": gt_entry.get("scenario_id", scenario_number), "error_type": None, "error_desc": None, "details": {}, "step_checks": {}, "eval_data": { "matches": eval_detail.get("tool_based", {}).get("matches", 0), "total_gt_calls": eval_detail.get("tool_based", {}).get("total_gt_calls", 0), "tool_based_success": eval_detail.get("tool_based", {}).get("success", False), "result_based_success": eval_detail.get("result_based", {}).get("success", False) } } interaction_calls_raw = interaction_entry.get("tool_calls", []) interaction_calls = extract_all_tool_calls(interaction_calls_raw) matches = eval_detail.get("tool_based", {}).get("matches", 0) total_gt_calls = eval_detail.get("tool_based", {}).get("total_gt_calls", 0) tool_based_success = eval_detail.get("tool_based", {}).get("success", False) result_based_success = eval_detail.get("result_based", {}).get("success", False) # First check if successful: correct if both tool_based and result_based succeed if tool_based_success and result_based_success: result["error_type"] = "correct" result["error_desc"] = "No error" result["details"]["correct"] = True # Still record check results as additional information has_syntax_error, syntax_desc = check_syntax_error(interaction_calls_raw) result["step_checks"]["syntax_error"] = has_syntax_error query_tools = SCENARIO_CONFIG[scenario_type]['query_tools'] has_multimodal_error, multimodal_desc = check_multimodal_error( gt_entry, interaction_calls, query_tools, scenario_type, scenario_number ) result["step_checks"]["multimodal_error"] = has_multimodal_error has_hallucination_error, hallucination_desc = check_hallucination_error(gt_entry, interaction_calls) result["step_checks"]["hallucination_error"] = has_hallucination_error if has_multimodal_error: result["details"]["multimodal_note"] = multimodal_desc if has_hallucination_error: result["details"]["hallucination_note"] = hallucination_desc result["step_checks"]["logic_error"] = False result["step_checks"]["memory_error"] = False result["step_checks"]["over_operation_error"] = False return result # Below: error reason determination for failed samples has_syntax_error, syntax_desc = check_syntax_error(interaction_calls_raw) result["step_checks"]["syntax_error"] = has_syntax_error if has_syntax_error: result["error_type"] = "syntax_error" result["error_desc"] = syntax_desc result["details"]["syntax_error"] = syntax_desc return result query_tools = SCENARIO_CONFIG[scenario_type]['query_tools'] has_multimodal_error, multimodal_desc = check_multimodal_error( gt_entry, interaction_calls, query_tools, scenario_type, scenario_number ) result["step_checks"]["multimodal_error"] = has_multimodal_error if has_multimodal_error: result["error_type"] = "multimodal_error" result["error_desc"] = multimodal_desc result["details"]["multimodal_error"] = multimodal_desc return result has_hallucination_error, hallucination_desc = check_hallucination_error(gt_entry, interaction_calls) result["step_checks"]["hallucination_error"] = has_hallucination_error if has_hallucination_error: result["error_type"] = "hallucination_error" result["error_desc"] = hallucination_desc result["details"]["hallucination_error"] = hallucination_desc return result diff = total_gt_calls - matches if diff >= 2: result["error_type"] = "logic_error" result["error_desc"] = f"Insufficient tool call matches: matches={matches}, total_gt_calls={total_gt_calls}, missing {diff}" result["details"]["logic_error"] = { "matches": matches, "total_gt_calls": total_gt_calls, "missing": diff } result["step_checks"]["logic_error"] = True result["step_checks"]["memory_error"] = False result["step_checks"]["over_operation_error"] = False return result if diff == 1: result["error_type"] = "memory_error" result["error_desc"] = f"Only the last tool call missed: matches={matches}, total_gt_calls={total_gt_calls}" result["details"]["memory_error"] = { "matches": matches, "total_gt_calls": total_gt_calls, "missing": 1 } result["step_checks"]["logic_error"] = False result["step_checks"]["memory_error"] = True result["step_checks"]["over_operation_error"] = False return result if matches == total_gt_calls and not result_based_success: result["error_type"] = "over_operation_error" result["error_desc"] = "Tool calls correct but result incorrect" result["details"]["over_operation_error"] = { "matches": matches, "total_gt_calls": total_gt_calls, "tool_based_success": tool_based_success, "result_based_success": result_based_success } result["step_checks"]["logic_error"] = False result["step_checks"]["memory_error"] = False result["step_checks"]["over_operation_error"] = True return result result["error_type"] = "unknown" result["error_desc"] = ( f"Unknown case: matches={matches}, total_gt_calls={total_gt_calls}, " f"tool_success={tool_based_success}, result_success={result_based_success}" ) return result def analyze_model_performance( model_name: str, scenario_type: str, scenario_number: int, mode: str, ground_truth: List[Dict], eval_result: Dict ) -> Dict: """Analyze model performance in a specific scenario""" if not is_usable_eval_result(eval_result): return { "error": f"Incomplete result file excluded: {scenario_type}{scenario_number}_{mode}.json" } interaction_log = load_interaction_log(model_name, scenario_type, scenario_number, mode) if interaction_log is None: return {"error": f"Result file not found: {scenario_type}{scenario_number}_{mode}.json"} results = [] for i, interaction_entry in enumerate(interaction_log): # Reconstructed result files may contain None placeholders for tasks that # were never (re)run; skip them — there is nothing to analyze. if interaction_entry is None: continue scenario_id = interaction_entry.get("scenario_id", i + 1) gt_entry = None for gt in ground_truth: if gt.get("scenario_id") == scenario_id: gt_entry = gt break if gt_entry is None: results.append({ "scenario_type": scenario_type, "scenario_number": scenario_number, "scenario_id": scenario_id, "error_type": "missing_gt", "error_desc": "Matching ground truth record not found" }) continue eval_detail = None for detail in eval_result.get("detailed_results", []): if detail.get("scenario_id") == scenario_id: eval_detail = detail break if eval_detail is None: results.append({ "scenario_type": scenario_type, "scenario_number": scenario_number, "scenario_id": scenario_id, "error_type": "missing_eval", "error_desc": "Matching evaluation record not found" }) continue results.append(analyze_scenario(scenario_type, gt_entry, interaction_entry, eval_detail, scenario_number)) return { "model": model_name, "scenario_type": scenario_type, "scenario": scenario_number, "mode": mode, "results": results } def calculate_error_rates(model_stats: Dict, model: str) -> Dict[str, float]: """ Calculate proportions of each error type 1. Syntax format error rate = syntax format error samples / total samples 2. Multimodal recognition error rate = multimodal error samples / syntax-correct samples 3. Hallucination error rate = incorrect user_id samples / syntax-correct samples 4. Logic reasoning error rate = logic error samples / samples without the first three errors 5. Memory/common sense error rate = memory error samples / samples without the first four errors 6. Over-operation error rate = over-operation error samples / tool-based correct samples """ stats = model_stats[model] total = stats["total"] if total == 0: return {} syntax_correct = total - stats["syntax_error"] no_multimodal_error = syntax_correct - stats["multimodal_error"] no_hallucination = no_multimodal_error - stats["hallucination_error"] no_logic_error = no_hallucination - stats["logic_error"] tool_correct = stats["correct"] + stats["over_operation_error"] rates = { "syntax_error_rate": stats["syntax_error"] / total if total > 0 else 0, "multimodal_error_rate": stats["multimodal_error"] / syntax_correct if syntax_correct > 0 else 0, "hallucination_rate": stats["hallucination_error"] / syntax_correct if syntax_correct > 0 else 0, "logic_reasoning_error_rate": stats["logic_error"] / no_hallucination if no_hallucination > 0 else 0, "memory_common_sense_error_rate": stats["memory_error"] / no_logic_error if no_logic_error > 0 else 0, "over_operation_error_rate": stats["over_operation_error"] / tool_correct if tool_correct > 0 else 0, "denominators": { "total": total, "syntax_correct": syntax_correct, "no_multimodal_error": no_multimodal_error, "no_hallucination": no_hallucination, "no_logic_error": no_logic_error, "tool_correct": tool_correct } } return rates def generate_report(all_results: List[Dict]) -> str: """Generate error analysis report""" report_lines = [] report_lines.append("=" * 120) report_lines.append("All-Scenario Error Reason Analysis Report (Based on eval_result Data)") report_lines.append("=" * 120) report_lines.append("") model_stats = defaultdict(lambda: { "total": 0, "syntax_error": 0, "multimodal_error": 0, "hallucination_error": 0, "logic_error": 0, "memory_error": 0, "over_operation_error": 0, "correct": 0 }) scenario_model_stats = defaultdict(lambda: defaultdict(lambda: defaultdict(lambda: { "total": 0, "syntax_error": 0, "multimodal_error": 0, "hallucination_error": 0, "logic_error": 0, "memory_error": 0, "over_operation_error": 0, "correct": 0 }))) for result in all_results: if "error" in result: continue model = result["model"] scenario_type = result["scenario_type"] scenario = result["scenario"] for scenario_result in result["results"]: if scenario_result.get("error_type") in ["missing_gt", "missing_eval"]: continue model_stats[model]["total"] += 1 scenario_model_stats[scenario_type][scenario][model]["total"] += 1 error_type = scenario_result.get("error_type") if error_type and error_type in model_stats[model]: model_stats[model][error_type] += 1 if error_type and error_type in scenario_model_stats[scenario_type][scenario][model]: scenario_model_stats[scenario_type][scenario][model][error_type] += 1 report_lines.append("I. Overall Error Type Distribution (by Model) - Raw Counts") report_lines.append("-" * 120) report_lines.append("") header = f"{'Model':<30} {'Total':>6} {'Syntax':>6} {'MultiMod':>8} {'Halluc':>8} {'Logic':>8} {'Memory':>8} {'OverOp':>10} {'Correct':>8}" report_lines.append(header) report_lines.append("-" * 120) for model in ACTIVE_MODELS: if model not in model_stats: continue stats = model_stats[model] total = stats["total"] if total == 0: continue row = f"{model:<30} {total:>6}" for error_type in [ "syntax_error", "multimodal_error", "hallucination_error", "logic_error", "memory_error", "over_operation_error", "correct" ]: row += f" {stats[error_type]:>8}" report_lines.append(row) report_lines.append("") report_lines.append("II. Error Type Rate Statistics (using defined denominators)") report_lines.append("-" * 120) report_lines.append("") for model in ACTIVE_MODELS: if model not in model_stats: continue stats = model_stats[model] if stats["total"] == 0: continue rates = calculate_error_rates(model_stats, model) denom = rates.get("denominators", {}) report_lines.append(f"\nModel: {model}") report_lines.append("-" * 80) report_lines.append(f"1. Syntax format error rate: {stats['syntax_error']}/{denom.get('total', 0)} = {rates.get('syntax_error_rate', 0):.2%}") if denom.get('syntax_correct', 0) > 0: report_lines.append(f"2. Multimodal recognition error rate: {stats['multimodal_error']}/{denom.get('syntax_correct', 0)} = {rates.get('multimodal_error_rate', 0):.2%}") else: report_lines.append("2. Multimodal recognition error rate: N/A (no syntax-correct samples)") if denom.get('syntax_correct', 0) > 0: report_lines.append(f"3. Hallucination (user_id) error rate: {stats['hallucination_error']}/{denom.get('syntax_correct', 0)} = {rates.get('hallucination_rate', 0):.2%}") else: report_lines.append("3. Hallucination (user_id) error rate: N/A (no syntax-correct samples)") no_hallucination = denom.get('no_hallucination', 0) if no_hallucination > 0: report_lines.append(f"4. Logic reasoning error rate: {stats['logic_error']}/{no_hallucination} = {rates.get('logic_reasoning_error_rate', 0):.2%}") else: report_lines.append("4. Logic reasoning error rate: N/A (no samples correct for first three checks)") no_logic_error = denom.get('no_logic_error', 0) if no_logic_error > 0: report_lines.append(f"5. Memory/common sense error rate: {stats['memory_error']}/{no_logic_error} = {rates.get('memory_common_sense_error_rate', 0):.2%}") else: report_lines.append("5. Memory/common sense error rate: N/A (no samples correct for first four checks)") tool_correct = denom.get('tool_correct', 0) if tool_correct > 0: report_lines.append(f"6. Over-operation error rate: {stats['over_operation_error']}/{tool_correct} = {rates.get('over_operation_error_rate', 0):.2%}") else: report_lines.append("6. Over-operation error rate: N/A (no tool-call-correct samples)") report_lines.append("") report_lines.append("III. Error Type Distribution by Scenario (by Model)") report_lines.append("-" * 120) for scenario_type in ALL_SCENARIO_TYPES: if scenario_type not in scenario_model_stats: continue report_lines.append("") report_lines.append(f"Scenario Type: {scenario_type}") report_lines.append("-" * 100) for scenario in sorted(scenario_model_stats[scenario_type].keys()): report_lines.append("") report_lines.append(f" Scenario {scenario}:") header = f"{'Model':<30} {'Total':>6} {'Syntax':>6} {'MultiM':>7} {'Halluc':>7} {'Logic':>7} {'Memory':>7} {'OverOp':>10} {'Correct':>7}" report_lines.append(header) for model in ACTIVE_MODELS: if model not in scenario_model_stats[scenario_type][scenario]: continue stats = scenario_model_stats[scenario_type][scenario][model] total = stats["total"] if total == 0: continue row = f"{model:<30} {total:>6}" for error_type in [ "syntax_error", "multimodal_error", "hallucination_error", "logic_error", "memory_error", "over_operation_error", "correct" ]: row += f" {stats[error_type]:>8}" report_lines.append(row) report_lines.append("") report_lines.append("IV. Error Type Summary Across All Models") report_lines.append("-" * 120) report_lines.append("") total_all = sum(stats["total"] for stats in model_stats.values()) error_type_totals = { "syntax_error": 0, "multimodal_error": 0, "hallucination_error": 0, "logic_error": 0, "memory_error": 0, "over_operation_error": 0, "correct": 0 } for stats in model_stats.values(): for error_type in error_type_totals: error_type_totals[error_type] += stats[error_type] syntax_correct_all = total_all - error_type_totals["syntax_error"] no_hallucination_all = ( syntax_correct_all - error_type_totals["multimodal_error"] - error_type_totals["hallucination_error"] ) no_logic_error_all = no_hallucination_all - error_type_totals["logic_error"] tool_correct_all = error_type_totals["correct"] + error_type_totals["over_operation_error"] report_lines.append("Error type counts and proportions:") report_lines.append("") report_lines.append( f"1. Syntax format error: {error_type_totals['syntax_error']}/{total_all} = {error_type_totals['syntax_error'] / total_all:.2%}" if total_all > 0 else "1. Syntax format error: N/A" ) if syntax_correct_all > 0: report_lines.append(f"2. Multimodal recognition error: {error_type_totals['multimodal_error']}/{syntax_correct_all} = {error_type_totals['multimodal_error'] / syntax_correct_all:.2%}") report_lines.append(f"3. Hallucination (user_id): {error_type_totals['hallucination_error']}/{syntax_correct_all} = {error_type_totals['hallucination_error'] / syntax_correct_all:.2%}") else: report_lines.append("2. Multimodal recognition error: N/A") report_lines.append("3. Hallucination (user_id): N/A") if no_hallucination_all > 0: report_lines.append(f"4. Logic reasoning error: {error_type_totals['logic_error']}/{no_hallucination_all} = {error_type_totals['logic_error'] / no_hallucination_all:.2%}") else: report_lines.append("4. Logic reasoning error: N/A") if no_logic_error_all > 0: report_lines.append(f"5. Memory/common sense error: {error_type_totals['memory_error']}/{no_logic_error_all} = {error_type_totals['memory_error'] / no_logic_error_all:.2%}") else: report_lines.append("5. Memory/common sense error: N/A") if tool_correct_all > 0: report_lines.append(f"6. Over-operation error: {error_type_totals['over_operation_error']}/{tool_correct_all} = {error_type_totals['over_operation_error'] / tool_correct_all:.2%}") else: report_lines.append("6. Over-operation error: N/A") report_lines.append("") report_lines.append( f"Correct samples: {error_type_totals['correct']}/{total_all} = {error_type_totals['correct'] / total_all:.2%}" if total_all > 0 else "Correct samples: N/A" ) report_lines.append("") report_lines.append(f"Total: {total_all} samples") report_lines.append("") report_lines.append("=" * 120) return "\n".join(report_lines) def main(): """Main function""" import argparse parser = argparse.ArgumentParser(description="All-scenario error reason analysis") parser.add_argument("--results_root", type=str, default=None, help="Root dir holding per-model interaction logs (default: results). " "Set to GPT_user_results to analyse the GPT-user run.") parser.add_argument("--eval_root", type=str, default=None, help="Root dir holding per-model eval output (default: eval_result). " "Set to GPT_user_eval_result to analyse the GPT-user run.") parser.add_argument("--output_root", type=str, default=None, help="Output dir for report + details JSON (default: error_analysis). " "Set to error_analysis_gpt_user to keep GPT-user results separate.") parser.add_argument("--models", type=str, default=None, help="Comma-separated model names to analyse (default: built-in MODELS list).") parser.add_argument( "--include_partial", action="store_true", help="Analyse every valid trajectory in incomplete evaluation files.", ) args = parser.parse_args() global RESULTS_ROOT, EVAL_ROOT, OUTPUT_ROOT, ACTIVE_MODELS, INCLUDE_PARTIAL if args.results_root is not None: RESULTS_ROOT = args.results_root if args.eval_root is not None: EVAL_ROOT = args.eval_root if args.output_root is not None: OUTPUT_ROOT = args.output_root if args.models is not None and args.models.strip(): ACTIVE_MODELS = [m.strip() for m in args.models.split(',') if m.strip()] INCLUDE_PARTIAL = args.include_partial print("Starting all-scenario error reason analysis...") print(f" results_root : {RESULTS_ROOT}") print(f" eval_root : {EVAL_ROOT}") print(f" output_root : {OUTPUT_ROOT}") print(f" models : {ACTIVE_MODELS}") print(f" include_partial: {INCLUDE_PARTIAL}") print("") all_results = [] for scenario_type in ALL_SCENARIO_TYPES: print(f"Processing scenario type {scenario_type}...") for scenario_number in SCENARIO_CONFIG[scenario_type]['scenario_numbers']: print(f" Processing scenario {scenario_number}...") ground_truth = load_ground_truth(scenario_type, scenario_number) if not ground_truth: print(" Skipped: unable to load ground truth file") continue for mode in MODES: for model in ACTIVE_MODELS: eval_result = load_eval_result(model, scenario_type, scenario_number, mode) if eval_result is None: continue if not is_usable_eval_result(eval_result): print(f" Skipped incomplete file: {model}/{scenario_type}{scenario_number}_{mode}.json") continue result = analyze_model_performance( model, scenario_type, scenario_number, mode, ground_truth, eval_result ) all_results.append(result) print("") print("Generating analysis report...") report = generate_report(all_results) print(report) output_dir = os.path.join(os.path.dirname(__file__), '..', OUTPUT_ROOT) os.makedirs(output_dir, exist_ok=True) output_path = os.path.join(output_dir, 'all_scenarios_error_analysis_report.txt') with open(output_path, 'w', encoding='utf-8') as f: f.write(report) print(f"\nReport saved to: {output_path}") json_path = os.path.join(output_dir, 'all_scenarios_error_analysis_details.json') with open(json_path, 'w', encoding='utf-8') as f: json.dump(all_results, f, ensure_ascii=False, indent=2) print(f"Detailed results saved to: {json_path}") if __name__ == "__main__": main()
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
All-Scenario Error Reason Analysis Script - Based on eval_result data
Supported scenarios: retail, restaurant (including multi-restaurant variant 6),
kitchen, warehouse, household
Error reasons are determined for each sample in the following order:
1. Syntax format error
2. Multimodal recognition error (whether query tool calls match key/value)
3. Hallucination issue (incorrect user_id parameter)
4. Tool call logic reasoning error (matches < total_gt_calls, difference >= 2)
5. Memory/common sense error (matches < total_gt_calls, difference == 1)
6. Over-operation error (matches == total_gt_calls but result_based failed)
Correct sample: both tool-based evaluation and result-based evaluation are correct
"""
import json
import os
import sys
from collections import defaultdict
from typing import Dict, List, Optional, Tuple
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
from tools.retail.retail_db import RetailDB
from tools.retail.retail_init import (
retail_init_data1, retail_init_data2, retail_init_data3, retail_init_data4,
retail_init_data5, retail_init_data6, retail_init_data7, retail_init_data8,
retail_init_data9, retail_init_data10
)
from tools.restaurant.restaurant_db import RestaurantDB
from tools.restaurant.restaurant_init import restaurant_init_data, restaurant_init_data5
from tools.restaurant.restaurant6_db import Restaurant6DB
from tools.restaurant.restaurant6_init import restaurant6_init_data
from tools.kitchen.kitchen_db import KitchenDB
from tools.kitchen.kitchen_init import kitchen_init_data
from tools.warehouse.warehouse_db import WarehouseDB
from tools.warehouse import warehouse_init
from tools.household.household_db import HouseholdDB
from tools.household import household_init
MODELS = [
'qwen3-vl-235b', 'Qwen3.5-397B-A17B', 'gemini-3.1-pro-preview',
'kimi-k2.6', 'mimo-v2.5-omni', 'qwen3.6-plus', 'glm-5v-turbo', 'doubao-seed-2-0-pro-260215'
]
# ===================== Configurable I/O Roots =====================
# Defaults preserve the original behaviour (reading results/ + eval_result/,
# writing error_analysis/). Override via CLI (see main()) to analyse a
# different run, e.g. the GPT-user run:
# --results_root GPT_user_results --eval_root GPT_user_eval_result
# --output_root error_analysis_gpt_user
RESULTS_ROOT = 'results'
EVAL_ROOT = 'eval_result'
OUTPUT_ROOT = 'error_analysis'
ACTIVE_MODELS = list(MODELS)
INCLUDE_PARTIAL = False
ERROR_TYPES = {
'syntax_error': 'Syntax format error in tool calls',
'multimodal_error': 'Multimodal recognition error',
'hallucination_error': 'Hallucination issue (incorrect user_id parameter)',
'logic_error': 'Tool call logic reasoning error',
'memory_error': 'Memory or common sense error',
'over_operation_error': 'Over-operation error',
'correct': 'Correct'
}
MODES = ['easy', 'hard', 'static']
SCENARIO_CONFIG = {
'retail': {
'scenario_numbers': list(range(1, 11)),
'init_data_map': {
1: retail_init_data1, 2: retail_init_data2, 3: retail_init_data3,
4: retail_init_data4, 5: retail_init_data5, 6: retail_init_data6,
7: retail_init_data7, 8: retail_init_data8, 9: retail_init_data9,
10: retail_init_data10
},
'db_class': RetailDB,
'query_tools': {
'get_price', 'get_tax_rate', 'get_category', 'get_discount', 'get_nutrition',
'find_products_by_nutritional_characteristic', 'find_products_by_taste',
'find_products_by_country_of_origin', 'find_products_by_price_range',
'list_discounted_products'
},
},
'restaurant': {
'scenario_numbers': list(range(1, 7)),
'init_data_map': {
1: restaurant_init_data, 2: restaurant_init_data,
3: restaurant_init_data, 4: restaurant_init_data,
5: restaurant_init_data5,
6: restaurant6_init_data,
},
'db_class': RestaurantDB,