Spaces:
Runtime error
Runtime error
File size: 11,974 Bytes
69ee6e4 | 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 | from __future__ import annotations
import os
import hashlib
import json
import random
import re
import time
from contextlib import contextmanager
from dataclasses import dataclass
from typing import Any
from llm_eval.benchmarks.mmlu_subset import MMLUSubsetDataset
from llm_eval.cache import ResponseCache
from llm_eval.config import RunConfig, build_run_manifest, load_env_file
from llm_eval.policy import merge_policy
from llm_eval.providers import InferenceRequest, build_provider_client
from llm_eval.providers.http import ProviderHTTPError
from llm_eval.storage import ArtifactStore
OPTION_RE = re.compile(r"\b([A-Z])\b")
@dataclass
class ExecutionSummary:
run_id: str
total_requests: int
total_errors: int
provider_metrics: dict[str, dict[str, Any]]
def _system_id(provider: str, model: str) -> str:
return f"{provider}:{model}"
def _finalize_metrics(provider_metrics: dict[str, dict[str, Any]]) -> None:
for metrics in provider_metrics.values():
attempted = metrics["attempted"]
metrics["accuracy"] = (metrics["correct"] / attempted) if attempted else 0.0
def _request_key(
*,
provider: str,
model: str,
sample_id: str,
prompt: str,
temperature: float,
max_tokens: int,
) -> str:
payload = {
"provider": provider,
"model": model,
"sample_id": sample_id,
"prompt": prompt,
"temperature": temperature,
"max_tokens": max_tokens,
}
return hashlib.sha256(json.dumps(payload, sort_keys=True).encode("utf-8")).hexdigest()[:24]
def _extract_option_letter(text: str) -> str | None:
match = OPTION_RE.search(text.upper())
return match.group(1) if match else None
def _correct_letter(answer_index: int) -> str:
return chr(65 + answer_index)
@contextmanager
def _temporary_env(overrides: dict[str, str] | None):
if not overrides:
yield
return
snapshot = {key: os.getenv(key) for key in overrides}
try:
for key, value in overrides.items():
os.environ[key] = value
yield
finally:
for key, old_value in snapshot.items():
if old_value is None:
os.environ.pop(key, None)
else:
os.environ[key] = old_value
def run_evaluation(
config: RunConfig,
policy_path: str = "configs/policy.yaml",
artifacts_root: str = "artifacts",
env_path: str = ".env",
env_overrides: dict[str, str] | None = None,
) -> ExecutionSummary:
load_env_file(env_path)
with _temporary_env(env_overrides):
merged_policy = merge_policy(config, policy_path=policy_path)
config.policy = merged_policy
manifest = build_run_manifest(config)
random.seed(config.seed)
store = ArtifactStore(artifacts_root=artifacts_root, run_id=manifest.run_id)
cache = ResponseCache(store.run_dir / "cache")
store.write_manifest(manifest.model_dump())
if config.benchmark.name != "mmlu_subset":
raise NotImplementedError("Current runner supports mmlu_subset only.")
dataset = MMLUSubsetDataset(
config.benchmark.dataset_path,
max_samples=config.benchmark.max_samples,
)
samples = list(dataset.load())
completed_keys = store.load_completed_keys()
provider_metrics: dict[str, dict[str, Any]] = {}
for provider in config.providers:
sid = _system_id(provider.provider, provider.model)
provider_metrics[sid] = {
"provider": provider.provider,
"model": provider.model,
"requests": 0,
"errors": 0,
"correct": 0,
"attempted": 0,
}
total_requests = 0
total_errors = 0
for provider_cfg in config.providers:
sid = _system_id(provider_cfg.provider, provider_cfg.model)
client = build_provider_client(
provider_config=provider_cfg,
timeout_seconds=config.policy.reliability.request_timeout_seconds,
)
for sample in samples:
prompt = sample.prompt()
req_key = _request_key(
provider=provider_cfg.provider,
model=provider_cfg.model,
sample_id=sample.sample_id,
prompt=prompt,
temperature=provider_cfg.temperature,
max_tokens=provider_cfg.max_tokens,
)
if req_key in completed_keys:
continue
provider_metrics[sid]["requests"] += 1
provider_metrics[sid]["attempted"] += 1
total_requests += 1
cached = cache.get(req_key)
if cached is not None:
response_text = str(cached["text"])
latency_ms = int(cached.get("latency_ms") or 0)
usage = cached.get("usage")
else:
attempt = 0
response_text = ""
latency_ms = 0
usage = None
while True:
attempt += 1
try:
response = client.generate(
InferenceRequest(
prompt=prompt,
temperature=provider_cfg.temperature,
max_tokens=provider_cfg.max_tokens,
)
)
response_text = response.text
latency_ms = response.latency_ms or 0
usage = response.usage
cache.set(
req_key,
{
"text": response_text,
"latency_ms": latency_ms,
"usage": usage,
},
)
break
except ProviderHTTPError as exc:
retryable = (
exc.status_code
in config.policy.reliability.retry.retryable_status_codes
)
if (
attempt < config.policy.reliability.retry.max_attempts
and retryable
):
backoff = config.policy.reliability.retry.backoff_seconds[
min(
attempt - 1,
len(config.policy.reliability.retry.backoff_seconds) - 1,
)
]
time.sleep(backoff)
continue
provider_metrics[sid]["errors"] += 1
total_errors += 1
store.append_error(
{
"run_id": manifest.run_id,
"provider": provider_cfg.provider,
"model": provider_cfg.model,
"sample_id": sample.sample_id,
"request_key": req_key,
"error_type": "ProviderHTTPError",
"error": str(exc),
"attempt": attempt,
}
)
break
except Exception as exc: # noqa: BLE001
provider_metrics[sid]["errors"] += 1
total_errors += 1
store.append_error(
{
"run_id": manifest.run_id,
"provider": provider_cfg.provider,
"model": provider_cfg.model,
"sample_id": sample.sample_id,
"request_key": req_key,
"error_type": type(exc).__name__,
"error": str(exc),
"attempt": attempt,
}
)
break
predicted = _extract_option_letter(response_text or "")
expected = _correct_letter(sample.answer_index)
is_correct = predicted == expected
if is_correct:
provider_metrics[sid]["correct"] += 1
store.append_result(
{
"run_id": manifest.run_id,
"system_id": sid,
"provider": provider_cfg.provider,
"model": provider_cfg.model,
"sample_id": sample.sample_id,
"category": sample.category,
"request_key": req_key,
"predicted": predicted,
"expected": expected,
"is_correct": is_correct,
"latency_ms": latency_ms,
"usage": usage,
"response_text": response_text,
}
)
requests_for_provider = provider_metrics[sid]["requests"]
errors_for_provider = provider_metrics[sid]["errors"]
if requests_for_provider > 0:
error_rate_percent = (errors_for_provider / requests_for_provider) * 100
if (
config.policy.budget.enforce_hard_stop
and requests_for_provider
>= config.policy.reliability.provider_error_rate_window_size_requests
and error_rate_percent
> config.policy.reliability.provider_error_rate_hard_stop_percent
):
_finalize_metrics(provider_metrics)
summary = ExecutionSummary(
run_id=manifest.run_id,
total_requests=total_requests,
total_errors=total_errors,
provider_metrics=provider_metrics,
)
store.write_summary(
{
"run_id": summary.run_id,
"status": "stopped_due_to_error_rate",
"total_requests": summary.total_requests,
"total_errors": summary.total_errors,
"provider_metrics": summary.provider_metrics,
}
)
return summary
_finalize_metrics(provider_metrics)
summary = ExecutionSummary(
run_id=manifest.run_id,
total_requests=total_requests,
total_errors=total_errors,
provider_metrics=provider_metrics,
)
store.write_summary(
{
"run_id": summary.run_id,
"status": "completed",
"total_requests": summary.total_requests,
"total_errors": summary.total_errors,
"provider_metrics": summary.provider_metrics,
}
)
return summary
|