EgoBench / analysis_scripts /evaluate_interaction_test.py
emodiary's picture
Initial complete EgoBench release
3c4b9d3 verified
Raw History Blame Contribute Delete
6.53 kB
import argparse
import json
import os
import tempfile
import unittest
from analysis_scripts.evaluate_interaction import (
evaluate_interaction_success,
get_init_db,
parse_result_filename,
)
class EvaluateNewScenariosTest(unittest.TestCase):
def test_restaurant6_is_classified_as_restaurant_variant_6(self):
self.assertEqual(
("restaurant", 6, "easy"),
parse_result_filename("restaurant6_easy.json"),
)
self.assertIsNone(parse_result_filename("restaurant7_easy.json"))
def test_household_result_filename_is_supported(self):
self.assertEqual(
("household", 18, "static"),
parse_result_filename("household18_static.json"),
)
self.assertIsNone(parse_result_filename("household19_static.json"))
def _evaluate(self, scenario, scenario_number, ground_truth, interactions):
with tempfile.TemporaryDirectory() as temp_dir:
ground_truth_path = os.path.join(temp_dir, "ground_truth.json")
interaction_path = os.path.join(temp_dir, "interaction.json")
with open(ground_truth_path, "w", encoding="utf-8") as stream:
json.dump(ground_truth, stream)
with open(interaction_path, "w", encoding="utf-8") as stream:
json.dump(interactions, stream)
return evaluate_interaction_success(
ground_truth_path,
interaction_path,
scenario=scenario,
args=argparse.Namespace(scenario_number=scenario_number),
silent=True,
)
def test_warehouse_mutation_and_final_state_are_evaluated(self):
database = get_init_db("warehouse", 1)
equipment = next(iter(database.equipment))
call = {
"tool_name": "add_to_equipment_action_list",
"parameters": {
"user_id": "warehouse_user_01",
"equipment": equipment,
"required_action": "replace",
"due_date": "2026-08-20",
},
}
result = self._evaluate(
"warehouse",
1,
[{"scenario_id": 1, "ground_truth": [call]}],
[{
"scenario_id": 1,
"tool_calls": [{"turn": 0, "calls": [call], "results": []}],
}],
)
self.assertTrue(result["is_complete"])
self.assertEqual(result["joint_success"]["success_rate"], 1.0)
def test_household_mutation_and_final_state_are_evaluated(self):
database = get_init_db("household", 1)
item_name = next(iter(database.items))
call = {
"tool_name": "add_to_household_action_list",
"parameters": {
"user_id": "household_user_01",
"name": item_name,
"required_action": "repair",
"due_date": "2026-08-20",
},
}
result = self._evaluate(
"household",
1,
[{"scenario_id": 1, "ground_truth": [call]}],
[{
"scenario_id": 1,
"tool_calls": [{"turn": 0, "calls": [call], "results": []}],
}],
)
self.assertTrue(result["is_complete"])
self.assertEqual(result["joint_success"]["success_rate"], 1.0)
def test_missing_scenario_id_marks_file_incomplete(self):
database = get_init_db("warehouse", 1)
equipment = next(iter(database.equipment))
call = {
"tool_name": "get_equipment_category",
"parameters": {"equipment": equipment},
}
result = self._evaluate(
"warehouse",
1,
[
{"scenario_id": 1, "ground_truth": [call]},
{"scenario_id": 2, "ground_truth": [call]},
],
[{
"scenario_id": 1,
"tool_calls": [{"turn": 0, "calls": [call], "results": []}],
}],
)
self.assertFalse(result["is_complete"])
self.assertEqual(result["valid_scenarios"], 1)
self.assertEqual(result["total_scenarios"], 2)
def test_order_is_not_a_supported_database_scenario(self):
with self.assertRaisesRegex(ValueError, "unsupported scenario"):
get_init_db("order", 1)
def test_only_four_original_simulated_user_averages_are_kept(self):
database = get_init_db("warehouse", 1)
equipment = next(iter(database.equipment))
call = {
"tool_name": "get_equipment_category",
"parameters": {"equipment": equipment},
}
ground_truth = [
{"scenario_id": 1, "ground_truth": [call]},
{"scenario_id": 2, "ground_truth": [call]},
]
interactions = [
{
"scenario_id": 1,
"tool_calls": [{"turn": 0, "calls": [call], "results": []}],
"user_performance": {
"original_role_consistency_avg": 1.0,
"original_instruction_following_avg": 0.8,
"original_resilience_avg": 0.6,
"original_contextual_robustness_avg": 0.4,
"final_role_consistency_avg": 99.0,
},
},
{
"scenario_id": 2,
"tool_calls": [{"turn": 0, "calls": [call], "results": []}],
"user_performance": {
"original_role_consistency_avg": 0.5,
},
},
]
result = self._evaluate(
"warehouse", 1, ground_truth, interactions
)
performance = result["performance_metrics"]
self.assertEqual(
set(performance["avg_user_performance"]),
{
"role_consistency",
"instruction_following",
"resilience",
"contextual_robustness",
},
)
self.assertAlmostEqual(
performance["avg_user_performance"]["role_consistency"], 0.75
)
self.assertEqual(
performance["user_performance_sample_counts"]["role_consistency"],
2,
)
self.assertEqual(
performance["user_performance_sample_counts"]["instruction_following"],
1,
)
self.assertNotIn("scenario_stats", result)
self.assertNotIn("filtered_user_issue", result)
if __name__ == "__main__":
unittest.main()