qqqwt's picture
Add files using upload-large-folder tool
b192407 verified
Raw
History Blame Contribute Delete
20.4 kB
"""
Evaluation Module - Evaluator for scoring agent outputs.
This module evaluates agent performance:
- Compares output files against gold standard
- Executes evaluation functions dynamically
- Calculates scores based on comparison results
- Tracks trajectory information (actions, token usage)
Reference: https://github.com/yiyihum/da-code/tree/main/da_agent/evaluators/evaluation.py
"""
import logging
import os, json
from typing import Callable, Any
from typing import List, Dict
from pathlib import Path
import sys, jsonlines
here=Path(__file__).absolute()
sys.path.append(str(here.parent))
from metrics import *
from tqdm import tqdm
from da_agent.envs.utils import timeout
import re
import traceback
Metric = Callable[[Any, Any], float]
class Evaluator:
def __init__(self, output_dir: str, gold_dir: str, timeout_seconds: int = 10):
self.output_dir = output_dir
self.gold_dir = gold_dir
self.timeout_second = timeout_seconds
def get_result_file(self, results: List, dir: str, isgold: bool):
results = results if isinstance(results, list)\
else [results]
if 'number' in results[0].keys():
return 'number', [results[0]['number']]
result_files = []
for result in results:
multi = result.get("multi", False)
files = result['file'] if isinstance(result['file'], list) \
else [result['file']]
if multi:
files = [os.path.join(dir, file) for file in files] if not isgold \
else [os.path.join(dir, os.path.basename(file)) for file in files]
result_files.append(files)
else:
for file in files:
file = file if not isgold else os.path.basename(file)
# if not os.path.exists(os.path.join(dir, file)):
# print(f"File not found : {os.path.join(dir, file)}")
result_files.append(os.path.join(dir, file))
return 'file', result_files
def _get_eval_config_info(self, eval_config: Dict[str, Any]):
# evaluator dict
# func -> metric function string, or list of metric function strings
# conj -> conjunction of multiple metrics if func is a list with length > 1, "and"/"or"
# result -> result getter config, or list of result getter configs
# expected (optional) -> expected getter config, or list of expected getter configs
# options (optional) -> metric options, or list of metric options
# if func is a str list, then result, expected (if exists), options (if exists) should also be lists of the same length
# even if one of the metrics does not need expected or options field, it should be included in the list with None
# self.evaluator = task_config["evaluator"]
id = eval_config['id']
output_id_dir = os.path.join(self.output_dir, id)
result_file = os.path.join(output_id_dir, 'dabench', 'result.json')
if not os.path.exists(result_file):
print(f"File {result_file} not found")
return id, False, None, None
trajectory_info = self._get_trajectory_info_from_json(result_file)
gold_id_dir = os.path.join(self.gold_dir, id)
if 'smolagents' in output_id_dir:
code_file = os.path.join(output_id_dir, 'code.ipynb')
code_content = {}
try:
with open(code_file, 'r', encoding='utf-8') as f:
import nbformat
notebook = nbformat.read(f, as_version=4)
cells = []
for idx, cell in enumerate(notebook.cells):
if cell.source and cell.source.strip():
cell_type = 'code' if cell.cell_type == 'code' else 'markdown'
cells.append({
'id': idx,
'type': cell_type,
'content': cell.source
})
code_content[os.path.basename(code_file)] = cells
except Exception as e:
logging.warning(f"Failed to read code file {code_file}: {e}")
if code_content:
trajectory_info['actions'] = code_content
config = {'domain': eval_config['domain'].split('/')[0]}
gold_file_name = eval_config['gold_file_name'] if isinstance(eval_config['gold_file_name'], list) \
else [eval_config['gold_file_name']]
gold_file_dict = {file_name: os.path.join(gold_id_dir, os.path.basename(file_name)) for file_name in gold_file_name}
output_file_name = eval_config['output_file_name'] if isinstance(eval_config['output_file_name'], list) \
else [eval_config['output_file_name']]
output_file_dict = {file_name: os.path.join(output_id_dir, os.path.basename(file_name)) for file_name in output_file_name}
eval_func = eval_config["eval_func"] \
if isinstance(eval_config["eval_func"], list)\
else [eval_config["eval_func"]]
exe_eval_func = []
for func in eval_func:
for k, v in gold_file_dict.items():
pattern = rf"(['\"]){k}\1"
matches = list(re.finditer(pattern, func))
if len(matches) > 1:
raise ValueError(f"Found multiple matches for {k}")
func = re.sub(pattern, rf"\1{v}\1", func)
for k, v in output_file_dict.items():
pattern = rf"(['\"]){k}\1"
matches = list(re.finditer(pattern, func))
if len(matches) > 1:
raise ValueError(f"Found multiple matches for {k}")
func = re.sub(rf"(['\"]){k}\1", rf"\1{v}\1", func)
exe_eval_func.append(func)
return id, True, trajectory_info, (config, exe_eval_func)
def _get_trajectory_info_from_json(self,result_file):
with open(result_file, 'r') as f:
result = json.load(f)
trajectory = result["trajectory"]
actions = []
total_token_usage = {'input_tokens': 0, 'output_tokens': 0, 'total_tokens': 0}
if 'smolagents' in result_file:
pass
elif 'da-agent' in result_file:
for i, step in enumerate(trajectory):
if i+1 < len(trajectory):
observation = trajectory[i+1]["observation"]
else:
observation = ""
actions.append({
"action": step["action"],
"content": step.get("code", "") if not step['action'].startswith('Python') else len(step.get("code", "").split('\n')),
"observation": observation,
"token_usage": step.get('usage'),
"timing": step.get('timing')
})
elif 'claude-code' in result_file:
num_turns = 0
duration_ms = 0
for i, step in enumerate(trajectory):
step_type = step.get("type", "")
if step_type == "result":
usage = step.get("usage", {})
if usage:
base_input = usage.get('input_tokens', 0)
cache_creation = usage.get('cache_creation_input_tokens', 0)
cache_read = usage.get('cache_read_input_tokens', 0)
output = usage.get('output_tokens', 0)
total_input = base_input + cache_creation + cache_read
total_token_usage = {
'input_tokens': total_input,
'output_tokens': output,
'cache_creation_input_tokens': cache_creation,
'cache_read_input_tokens': cache_read,
'total_tokens': total_input + output
}
cache_detail = usage.get('cache_creation', {})
if cache_detail:
total_token_usage['cache_creation_5m_input_tokens'] = cache_detail.get('ephemeral_5m_input_tokens', 0)
total_token_usage['cache_creation_1h_input_tokens'] = cache_detail.get('ephemeral_1h_input_tokens', 0)
num_turns = step.get("num_turns", 0)
duration_ms = step.get("duration_ms", 0)
continue
if step_type in ("system", "system_init"):
continue
if step_type == "assistant":
content = step.get("content", "")
tool_uses = step.get("tool_uses", [])
# Record text content if non-empty
if content:
actions.append({
"action": "Think",
"content": content,
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
# Record each tool use separately
for tool_use in tool_uses:
tool_name = tool_use.get("name", "")
tool_input = tool_use.get("input", {})
if tool_name == "Bash":
action_content = tool_input.get("command", "")
elif tool_name in ("Write", "Edit"):
action_content = tool_input.get("file_path", "")
else:
action_content = str(tool_input)
actions.append({
"action": tool_name,
"content": action_content,
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
elif step_type == "user":
# Tool result: extract content from message.content[].content
message = step.get("message", {})
content_list = message.get("content", [])
result_texts = []
for item in (content_list if isinstance(content_list, list) else []):
if isinstance(item, dict) and item.get("type") == "tool_result":
c = item.get("content", "")
result_texts.append(str(c) if not isinstance(c, str) else c)
actions.append({
"action": "Observation",
"content": "\n".join(result_texts) if result_texts else str(content_list),
"tool_use_result": step.get("tool_use_result"),
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
else:
actions.append({
"action": step_type,
"content": str(step.get("content", "")),
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
elif 'codex' in result_file:
for step in trajectory:
step_type = step.get("type", "")
if step_type == "result":
usage = step.get("usage", {})
if usage:
input_tokens = usage.get('input_tokens', 0)
cached_tokens = usage.get('cached_input_tokens', 0)
output_tokens = usage.get('output_tokens', 0)
total_token_usage = {
'input_tokens': input_tokens + cached_tokens,
'output_tokens': output_tokens,
'cached_input_tokens': cached_tokens,
'total_tokens': input_tokens + cached_tokens + output_tokens
}
continue
if step_type in ("thread_started", "turn_started", "item_started", "item_updated"):
continue
if step_type == "assistant":
content = step.get("content", "")
code_action = step.get("code_action")
observations = step.get("observations", "")
if code_action:
action = {
"action": "Bash",
"content": code_action,
"observation": observations,
"token_usage": step.get("usage"),
"timing": step.get('timing')
}
exit_code = step.get("exit_code")
if exit_code is not None:
action["exit_code"] = exit_code
actions.append(action)
elif content:
actions.append({
"action": "Think",
"content": content,
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
elif step_type == "error":
error_msg = step.get("content", "")
actions.append({
"action": "Error",
"content": error_msg,
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
elif step_type == "turn.failed":
error_detail = step.get("content", "")
actions.append({
"action": "TurnFailed",
"content": error_detail,
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
else:
actions.append({
"action": step_type,
"content": str(step.get("content", "")),
"token_usage": step.get("usage"),
"timing": step.get('timing')
})
info = {"finished": result["finished"], "steps": result["steps"],
"result": result["result"],
"added_files": result["result_files"]["added_files"],
"changed_files": result["result_files"]["changed_files"],
"actions": actions}
if 'claude-code' in result_file:
info['total_token_usage'] = total_token_usage
info['num_turns'] = num_turns
info['duration_ms'] = duration_ms
elif 'codex' in result_file:
info['total_token_usage'] = total_token_usage
return info
def _get_result_file_from_json(self, output_id_dir, result_file, is_plot=False):
pattern = r'\b(?:[\w/\-_]+/)?([\w\-_]+(\.\w+)+)\b'
filenames = re.findall(pattern, result_file)
if not filenames:
return []
filenames = [filename[0] for filename in filenames]
result_file = [os.path.join(output_id_dir, file) for file in filenames]
if is_plot:
result_file += [os.path.join(output_id_dir,"dabench/plot.json"), os.path.join(output_id_dir,"dabench/result.npy")]
result_file = [result_file]
return result_file
def evaluate(self, env_config: Dict|str):
"""
Evaluate task
"""
if isinstance(env_config, str):
if not os.path.exists(env_config) or not os.path.isfile(env_config):
raise ValueError('File Path Error: Please provide a right file path')
if env_config.endswith('.json'):
with open(env_config, 'r') as f:
env_configs = json.load(f)
elif env_config.endswith('.jsonl'):
with jsonlines.open(env_config, 'r') as js:
env_configs = [config_eval for config_eval in js]
else:
raise ValueError('File Type Error: Please Upload json or jsonl file')
env_configs = env_configs if isinstance(env_configs, list) else [env_configs]
elif isinstance(env_config, dict):
env_configs = [env_config]
eval_results = []
pbar = tqdm(total=len(env_configs))
for eval_config in env_configs:
id, exist, trajectory_info, eval_info = self._get_eval_config_info(eval_config)
pbar.set_description(f"Processing Task id: {id}")
pbar.update(1)
if not exist:
print(f"Result of Task {id} does not exist!")
continue
(config, eval_func) = eval_info
domain = config['domain']
try:
with timeout(self.timeout_second,"Action execution time exceeded!"):
scores = []
info = []
for func in eval_func:
# Capture logging.warning
class WarningCaptureHandler(logging.Handler):
def __init__(self):
super().__init__()
self.warnings = []
def emit(self, record):
if record.levelno >= logging.WARNING:
self.warnings.append(self.format(record))
# Create handler and add to root logger
warning_handler = WarningCaptureHandler()
warning_handler.setLevel(logging.WARNING)
formatter = logging.Formatter('%(levelname)s - %(message)s')
warning_handler.setFormatter(formatter)
root_logger = logging.getLogger()
original_level = root_logger.level
root_logger.addHandler(warning_handler)
root_logger.setLevel(min(logging.WARNING, original_level))
try:
result = eval(func)
except FileNotFoundError as e:
logging.error(f"File not found! Error: {e}")
scores.append(0.0)
continue
finally:
# Remove handler and restore
root_logger.removeHandler(warning_handler)
root_logger.setLevel(original_level)
if isinstance(result, dict):
scores.append(result.get('score', 0.0))
# Add captured warnings to result
# if warning_handler.warnings:
# result['warnings'] = warning_handler.warnings
info.append(result)
else:
scores.append(result)
# For non-dict results, we still want to capture warnings
if warning_handler.warnings:
result_dict = {
'score': result,
# 'warnings': warning_handler.warnings
}
info.append(result_dict)
except Exception as e:
logging.error(f"Error in task {id}: {e}")
traceback.print_exc()
scores.append(0.0)
info.append({"score": 0.0, "errors": [str(e)]})
scores = [score if isinstance(score, float) or isinstance(score, int) else 0.0 for score in scores]
total_score = sum(scores) / len(scores)
eval_results.append({"id": id, "total_score": total_score, **trajectory_info,
'info': info, 'domain': domain})
return eval_results