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 | class BaseBenchmark:
"""Base class for benchmarks.
Evaluation phases:
- preprocess: load and preprocess the data
- rollout: rollout the predictions
- judge: judge the correctness of a batch of predictions
- stat: get metrics.
"""
dataset: DBDataManager
_source_to_processer: dict[str, BaseProcesser] = {}
_source_to_agent: dict[str, BaseAgent] = {}
def __init__(self, config: EvalConfig | str) -> None:
# config
if isinstance(config, str):
config = ConfigLoader.load_eval_config(name=config)
self.config = config
# dataset
self.dataset = DBDataManager(config)
_samples = self.dataset.load()
if len(_samples) == 0:
raise ValueError(f"No samples found for data config '{self.config.data}'! Please check the data config.")
async def main(self):
logger.info(f"> Running with config: \n{json.dumps(self.config.model_dump(), indent=2, ensure_ascii=False)}")
self.preprocess()
await self.rollout()
await self.judge()
logger.info("> Running stat...")
await self.stat()
logger.info("> Cleaning up...")
await self.cleanup()
def preprocess(self) -> None:
"""Preprocess the dataset before rollout."""
samples = self.dataset.get_samples(stage="init")
logger.info(f"Preprocessing {len(samples)} samples...")
results = []
for sample in tqdm(samples, desc="Preprocessing"):
processed_sample = self.preprocess_one(sample)
if processed_sample is not None:
results.append(processed_sample)
logger.info(f"Successfully preprocessed {len(results)} samples. Updated to db.")
return results
def preprocess_one(self, sample: EvaluationSample) -> EvaluationSample:
processer = self._get_processer(sample.source)
processed_sample = processer.preprocess_one(sample)
if processed_sample is None:
return None
self.dataset.save(sample)
return sample
async def rollout(self) -> None:
"""Rollout the datapoints."""
samples = self.dataset.get_samples(stage="init")
logger.info(f"Rollout {len(samples)} samples...")
semaphore = asyncio.Semaphore(self.config.concurrency)
async def rollout_with_semaphore(item: EvaluationSample):
async with semaphore:
try:
return await self.rollout_one(item)
except Exception as e: # pylint: disable=broad-except
logger.error(
f">>>>>>>>>>>>>\nError running rollout on sample '{item.raw_question}': {e}\n<<<<<<<<<<<<<",
exc_info=True,
)
tasks = [rollout_with_semaphore(item) for item in samples]
results = []
for task in tqdm(asyncio.as_completed(tasks), total=len(tasks), desc="Rolling out"):
result = await task
if result is not None:
results.append(result)
logger.info(f"Successfully rollout {len(results)} samples. Updated to db.")
return results
async def rollout_one(self, sample: EvaluationSample) -> EvaluationSample:
agent = get_agent(self.config.agent)
await agent.build()
trace_id = AgentsUtils.gen_trace_id()
start_time = time.time()
result = await agent.run(sample.augmented_question, trace_id=trace_id)
end_time = time.time()
# Update the sample with the predicted answer and trajectory
sample.update(
trace_id=trace_id,
response=result.final_output,
time_cost=end_time - start_time,
trajectories=json.dumps(result.trajectories, ensure_ascii=False),
stage="rollout", # update stage to rollout!
)
self.dataset.save(sample)
return sample
async def judge(self, stage: str | None = "rollout") -> list[EvaluationSample]:
"""Judge samples.
Args:
stage (str|None, optional): The stage of samples to judge. If set to None, you can rejudge all samples.
"""
samples = self.dataset.get_samples(stage=stage)
logger.info(f"Judging {len(samples)} samples...")
semaphore = asyncio.Semaphore(self.config.judge_concurrency)
async def judge_with_semaphore(item: EvaluationSample):
async with semaphore:
try:
return await self.judge_one(item)
except Exception as e: # pylint: disable=broad-except
logger.error(f">>>>>>>>>>>>>\nError judging sample '{item}': {e}\n<<<<<<<<<<<<<", exc_info=True)
return None
tasks = [judge_with_semaphore(item) for item in samples]
results = []
for task in tqdm(asyncio.as_completed(tasks), total=len(tasks), desc="Judging"):
result = await task
if result is not None:
results.append(result)
logger.info(f"Successfully judged {len(results)} samples. Updated to db.")
return results
async def judge_one(self, data: EvaluationSample) -> EvaluationSample:
judger = self._get_processer(data.source)
result = await judger.judge_one(data)
result.update(stage="judged") # update stage to judged
self.dataset.save(result)
return result
async def stat(self) -> list[dict]:
# TODO: wrap the data like @verl / @torch
# TODO: log to wandb
judged_samples = self.dataset.get_samples(stage="judged")
logger.info(f"Stat from {len(judged_samples)} samples:")
data_by_benchmark = self._group_data_by_benchmark(judged_samples)
overall_results: list[dict] = []
for benchmark, data in data_by_benchmark.items():
evaluator = self._get_processer(benchmark)
result = await evaluator.stat(data)
overall_results.append(result)
logger.info(json.dumps(overall_results, indent=4, ensure_ascii=False))
return overall_results
def _get_processer(self, source: str) -> BaseProcesser:
if source not in self._source_to_processer:
processer = PROCESSER_FACTORY.get(source, self.config)
self._source_to_processer[source] = processer
return self._source_to_processer[source]
def _group_data_by_benchmark(self, predict_data: list[EvaluationSample]) -> dict[str, list[EvaluationSample]]:
# group data by benchmark
data_by_benchmark: dict[str, list[EvaluationSample]] = {}
for data in predict_data:
benchmark = data.source
if benchmark not in data_by_benchmark:
data_by_benchmark[benchmark] = []
data_by_benchmark[benchmark].append(data)
return data_by_benchmark
async def cleanup(self):
for agent in self._source_to_agent.values():
await agent.cleanup()
|