File size: 12,060 Bytes
246d201 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 |
import asyncio
import os
import re
import tempfile
from typing import Any
import pandas as pd
from datasets import load_dataset
from evaluation.benchmarks.agent_bench.helper import (
FAKE_RESPONSES,
INST_SUFFIXES,
compare_results,
create_sh_file,
)
from evaluation.utils.shared import (
EvalMetadata,
EvalOutput,
compatibility_for_eval_history_pairs,
make_metadata,
prepare_dataset,
reset_logger_for_multiprocessing,
run_evaluation,
)
from openhands.controller.state.state import State
from openhands.core.config import (
AppConfig,
SandboxConfig,
get_llm_config_arg,
parse_arguments,
)
from openhands.core.logger import openhands_logger as logger
from openhands.core.main import create_runtime, run_controller
from openhands.events.action import AgentFinishAction, CmdRunAction, MessageAction
from openhands.events.observation import CmdOutputObservation
from openhands.runtime.base import Runtime
from openhands.utils.async_utils import call_async_from_sync
def get_config(
metadata: EvalMetadata,
) -> AppConfig:
config = AppConfig(
default_agent=metadata.agent_class,
run_as_openhands=False,
runtime=os.environ.get('RUNTIME', 'docker'),
max_iterations=metadata.max_iterations,
sandbox=SandboxConfig(
base_container_image='python:3.12-slim',
enable_auto_lint=True,
use_host_network=False,
api_key=os.environ.get('ALLHANDS_API_KEY', None),
remote_runtime_api_url=os.environ.get('SANDBOX_REMOTE_RUNTIME_API_URL'),
keep_runtime_alive=False,
remote_runtime_init_timeout=3600,
),
# do not mount workspace
workspace_base=None,
workspace_mount_path=None,
)
config.set_llm_config(metadata.llm_config)
agent_config = config.get_agent_config(metadata.agent_class)
agent_config.enable_prompt_extensions = False
return config
def initialize_runtime(
runtime: Runtime,
instance: pd.Series, # this argument is not required
):
"""Initialize the runtime for the agent.
This function is called before the runtime is used to run the agent.
"""
logger.info(f"{'-' * 50} BEGIN Runtime Initialization Fn {'-' * 50}")
obs: CmdOutputObservation
# Set instance id
action = CmdRunAction(command='mkdir -p /workspace')
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
assert obs.exit_code == 0
action = CmdRunAction(command='cd /workspace')
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
assert obs.exit_code == 0
init_cmd = instance.init
if init_cmd is not None:
script_name = f'{instance.instance_id}_init.sh'
with tempfile.TemporaryDirectory() as tmpdir:
host_script_path = os.path.join(tmpdir, script_name)
create_sh_file(host_script_path, init_cmd)
runtime.copy_to(
host_script_path,
'/workspace',
)
logger.info(f'Running init script: {script_name}')
action = CmdRunAction(command=f'chmod +x ./{script_name} && ./{script_name}')
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
logger.info(obs, extra={'msg_type': 'OBSERVATION'})
assert obs.exit_code == 0
logger.info(f"{'-' * 50} END Runtime Initialization Fn {'-' * 50}")
def complete_runtime(
runtime: Runtime,
instance: pd.Series, # this argument is not required, but it is used to get the workspace_dir_name
) -> dict[str, Any]:
"""Complete the runtime for the agent.
This function is called before the runtime is used to run the agent.
If you need to do something in the sandbox to get the correctness metric after
the agent has run, modify this function.
"""
logger.info(f"{'-' * 50} BEGIN Runtime Completion Fn {'-' * 50}")
obs: CmdOutputObservation
agent_answer = None
get_agent_result_cmd = instance.get_agent_result
if get_agent_result_cmd is not None:
script_name = 'get_agent_result.sh'
with tempfile.TemporaryDirectory() as tmpdir:
host_script_path = os.path.join(tmpdir, script_name)
create_sh_file(host_script_path, get_agent_result_cmd)
runtime.copy_to(
host_script_path,
'/workspace',
)
logger.info(f'Running get agent result cmd: {script_name}')
action = CmdRunAction(
command=f'chmod +x ./{script_name} && ./{script_name}',
)
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
logger.info(obs, extra={'msg_type': 'OBSERVATION'})
assert obs.exit_code == 0
agent_answer = obs.content
# IF the agent answer is not found, retrieve it from the history
# We wait until the controller finishes
final_ans = None
if instance.ground_truth is not None:
final_ans = instance.ground_truth
else:
get_ground_truth_cmd = instance.get_ground_truth
if get_ground_truth_cmd is not None:
script_name = 'get_ground_truth.sh'
with tempfile.TemporaryDirectory() as tmpdir:
host_script_path = os.path.join(tmpdir, script_name)
create_sh_file(host_script_path, get_ground_truth_cmd)
runtime.copy_to(
host_script_path,
'/workspace',
)
logger.info(f'Running get ground truth cmd: {script_name}')
action = CmdRunAction(
command=f'chmod +x ./{script_name} && ./{script_name}'
)
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
logger.info(obs, extra={'msg_type': 'OBSERVATION'})
final_ans = obs.content
logger.info(f"{'-' * 50} END Runtime Completion Fn {'-' * 50}")
return {
'final_ans': final_ans,
'agent_answer': agent_answer,
}
def process_instance(
instance: pd.Series,
metadata: EvalMetadata,
reset_logger: bool = True,
) -> EvalOutput:
config = get_config(metadata)
# Setup the logger properly, so you can run multi-processing to parallelize the evaluation
if reset_logger:
log_dir = os.path.join(metadata.eval_output_dir, 'infer_logs')
reset_logger_for_multiprocessing(logger, instance.instance_id, log_dir)
else:
logger.info(f'Starting evaluation for instance {instance.instance_id}.')
# =============================================
# build instruction
# =============================================
# Prepare instruction
instruction = (
f'Please fix the following issue.\n'
'IMPORTANT: You should ONLY interact with the environment provided to you AND NEVER ASK FOR HUMAN HELP.\n'
'Please encapsulate your final answer (answer ONLY) within <solution> and </solution>.\n'
'For example: The answer to the question is <solution> 42 </solution>.\n'
'# Problem \n'
f'{instance.description}\n\n'
)
instruction += (
'IMPORTANT: You should ONLY interact with the environment provided '
'to you AND NEVER ASK FOR HUMAN HELP.\n'
)
# NOTE: You can actually set slightly different instruction for different agents
instruction += INST_SUFFIXES[metadata.agent_class]
# =============================================
# create sandbox and run the agent
# =============================================
runtime: Runtime = create_runtime(config)
call_async_from_sync(runtime.connect)
initialize_runtime(runtime, instance=instance)
# Here's how you can run the agent (similar to the `main` function) and get the final task state
state: State | None = asyncio.run(
run_controller(
config=config,
initial_user_action=MessageAction(content=instruction),
runtime=runtime,
fake_user_response_fn=FAKE_RESPONSES[metadata.agent_class],
)
)
if state is None:
raise ValueError('State should not be None.')
# =============================================
# result evaluation
# =============================================
return_val = complete_runtime(runtime, instance)
agent_answer = return_val['agent_answer']
final_ans = return_val['final_ans']
# If the agent answer is not found, retrieve it from the history
if agent_answer is None:
agent_answer = ''
logger.info('Retrieving agent answer from history.')
raw_ans = ''
# retrieve the last agent message or thought
for event in reversed(state.history):
if event.source == 'agent':
if isinstance(event, AgentFinishAction):
raw_ans = event.thought
break
elif isinstance(event, MessageAction):
raw_ans = event.content
break
elif isinstance(event, CmdRunAction):
raw_ans = event.thought
break
# parse the answer for a solution tag
agent_answer = re.findall(r'<solution>(.*?)</solution>', raw_ans, re.DOTALL)
if len(agent_answer) == 0:
logger.warning(f'Failed to parse model answer: {raw_ans}')
agent_answer = raw_ans
else:
agent_answer = agent_answer[0]
comparison_method = instance.comparison_method
logger.info(
f'Final message: {agent_answer} | Ground truth: {final_ans} | Comparison method: {comparison_method}'
)
test_result = compare_results(comparison_method, agent_answer, final_ans)
# history is now available as a stream of events, rather than list of pairs of (Action, Observation)
# for compatibility with the existing output format, we can remake the pairs here
# remove when it becomes unnecessary
histories = compatibility_for_eval_history_pairs(state.history)
metrics = state.metrics.get() if state.metrics else None
# Save the output
output = EvalOutput(
instance_id=instance.instance_id,
instance=instance.to_dict(),
instruction=instruction,
metadata=metadata,
history=histories,
metrics=metrics,
error=state.last_error if state and state.last_error else None,
test_result={
'agent_answer': agent_answer,
'final_answer': final_ans,
'check_method': comparison_method,
'result': test_result,
},
)
return output
if __name__ == '__main__':
args = parse_arguments()
dataset = load_dataset('iFurySt/AgentBench')
agent_bench_tests = dataset['osbench'].to_pandas()
llm_config = None
if args.llm_config:
llm_config = get_llm_config_arg(args.llm_config)
# modify_params must be False for evaluation purpose, for reproducibility and accurancy of results
llm_config.modify_params = False
if llm_config is None:
raise ValueError(f'Could not find LLM config: --llm_config {args.llm_config}')
metadata = make_metadata(
llm_config,
'AgentBench-OS',
args.agent_cls,
args.max_iterations,
args.eval_note,
args.eval_output_dir,
)
output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
instances = prepare_dataset(agent_bench_tests, output_file, args.eval_n_limit)
run_evaluation(
instances, metadata, output_file, args.eval_num_workers, process_instance
)
|