File size: 6,131 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
import asyncio
import os
import re

import nltk
import pandas as pd
from datasets import load_dataset

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 MessageAction

# Only CodeActAgent can delegate to BrowsingAgent
SUPPORTED_AGENT_CLS = {'CodeActAgent'}


def get_config(

    metadata: EvalMetadata,

) -> AppConfig:
    assert (
        metadata.max_iterations == 1
    ), 'max_iterations must be 1 for browsing delegation evaluation.'
    config = AppConfig(
        default_agent=metadata.agent_class,
        run_as_openhands=False,
        runtime='docker',
        max_iterations=metadata.max_iterations,
        sandbox=SandboxConfig(
            base_container_image='python:3.12-bookworm',
            enable_auto_lint=False,
            use_host_network=False,
        ),
        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 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}.')

    instruction = (
        f'You can delegate browsing tasks to a browser agent. '
        f"For example, for query 'Who is the president of the United States?', you can delegate the task to a browser agent via <execute_browse> Who is the president of the United States? </execute_browse>.\n"
        f'Now, solve the following query: "{instance.instruction}"\n'
        f'NOTE: You should copy the "query" as is into the <execute_browse> tag. DO NOT change ANYTHING in the query.'
    )

    runtime = create_runtime(config)

    state: State | None = asyncio.run(
        run_controller(
            config=config,
            initial_user_action=MessageAction(content=instruction),
            runtime=runtime,
        )
    )

    if state is None:
        raise ValueError('State should not be None.')

    metrics = state.metrics.get() if state.metrics else None
    # 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)

    # find the last delegate action
    last_delegate_action = None
    result = {}
    for action, _ in histories:
        if action['action'] == 'delegate':
            last_delegate_action = action
            instruction_for_delegate = action['args']['inputs']['task']
            # parse `browse_actions` from `instruction_for_delegate`
            # task = f'{thought}. I should start with: {browse_actions}'
            instruction_for_delegate = re.search(
                r'I should start with: (.*)', instruction_for_delegate
            ).group(1)

            # calculate the edit distance between the instance.instruction and the instruction_for_delegate
            edit_distance = nltk.edit_distance(
                instance.instruction, instruction_for_delegate
            )
            is_exact_match = (
                instance.instruction.strip() == instruction_for_delegate.strip()
            )
            result['edit_distance'] = edit_distance
            result['is_exact_match'] = is_exact_match

    # Save the output
    output = EvalOutput(
        instance_id=instance.instance_id,
        instruction=instruction,
        metadata=metadata,
        history=histories,
        metrics=metrics,
        error=state.last_error if state and state.last_error else None,
        test_result={
            'query': instance.instruction,
            'action': last_delegate_action,
            'result': result,
        },
    )
    return output


if __name__ == '__main__':
    args = parse_arguments()

    dataset = load_dataset('OpenHands/eval-browsing-instructions')
    dataset = dataset['train'].to_pandas()
    assert dataset.columns.tolist() == ['instance_id', 'instruction']

    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,
        'browsing_delegation',
        args.agent_cls,
        args.max_iterations,
        args.eval_note,
        args.eval_output_dir,
    )

    if metadata.agent_class not in SUPPORTED_AGENT_CLS:
        raise ValueError(
            f'Agent class {metadata.agent_class} not supported with AgentDelegation.'
        )

    output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
    instances = prepare_dataset(dataset, output_file, args.eval_n_limit)
    run_evaluation(
        instances,
        metadata,
        output_file,
        args.eval_num_workers,
        process_instance,
    )