Spaces:
Sleeping
Sleeping
File size: 15,972 Bytes
0b0338d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 | # server/benchmark_runner.py
"""
Benchmark Runner + Leaderboard β v4.0
Automatically runs ALL tasks Γ selected agent configurations and generates
a research-grade leaderboard output with per-task, per-strategy breakdowns.
Unlike existing benchmarks (SWE-bench, HumanEval) which require manual setup,
this runs end-to-end in-process with deterministic strategies.
Output format:
- Leaderboard table (ranked by composite score)
- Per-task breakdown
- Per-failure-type breakdown
- Generalization score (variance across tasks)
- Robustness score (from counterfactual engine)
- A "benchmark JSON" suitable for publishing or comparing systems
"""
from __future__ import annotations
import time
import json
from typing import List, Dict, Any, Optional
from dataclasses import dataclass, field
@dataclass
class BenchmarkResult:
"""Result of running one agent on one task variant."""
agent_name: str
task: str
variant_id: str
final_score: float
total_steps: int
cumulative_reward: float
duration_seconds: float
strategy: str
failure_type: str
reliability_index: float
causal_score: float
robustness_score: float
calibration_score: float
action_sequence: List[str]
@dataclass
class AgentBenchmarkSummary:
"""Aggregated results for one agent across all tasks."""
agent_name: str
tasks_run: int
mean_score: float
std_score: float
generalization_score: float # 1 - std (lower variance = more generalizable)
mean_steps: float
best_task: str
worst_task: str
mean_reliability: float
mean_causal_score: float
mean_robustness_score: float
mean_calibration_score: float
dominant_strategy: str
dominant_failure: str
composite_rank_score: float # Weighted final score for leaderboard
per_task_scores: Dict[str, float]
def to_dict(self) -> dict:
return {
"agent_name": self.agent_name,
"tasks_run": self.tasks_run,
"scores": {
"mean": round(self.mean_score, 3),
"std": round(self.std_score, 3),
"generalization": round(self.generalization_score, 3),
"reliability": round(self.mean_reliability, 3),
"causal_reasoning": round(self.mean_causal_score, 3),
"robustness": round(self.mean_robustness_score, 3),
"calibration": round(self.mean_calibration_score, 3),
"composite": round(self.composite_rank_score, 3),
},
"efficiency": {
"mean_steps": round(self.mean_steps, 1),
},
"behavior": {
"dominant_strategy": self.dominant_strategy,
"dominant_failure": self.dominant_failure,
},
"per_task_scores": {k: round(v, 3) for k, v in self.per_task_scores.items()},
"best_task": self.best_task,
"worst_task": self.worst_task,
}
@dataclass
class LeaderboardReport:
"""Full benchmark leaderboard."""
benchmark_id: str
tasks_evaluated: List[str]
agents_evaluated: List[str]
total_episodes: int
run_duration_seconds: float
rankings: List[AgentBenchmarkSummary]
raw_results: List[BenchmarkResult]
def to_dict(self) -> dict:
return {
"benchmark_id": self.benchmark_id,
"tasks_evaluated": self.tasks_evaluated,
"agents_evaluated": self.agents_evaluated,
"total_episodes": self.total_episodes,
"run_duration_seconds": round(self.run_duration_seconds, 2),
"leaderboard": [r.to_dict() for r in self.rankings],
"winner": self.rankings[0].agent_name if self.rankings else "none",
"insights": self._generate_insights(),
}
def _generate_insights(self) -> List[str]:
if not self.rankings:
return []
insights = []
top = self.rankings[0]
bottom = self.rankings[-1]
if top.composite_rank_score - bottom.composite_rank_score > 0.2:
insights.append(
f"Large performance gap: '{top.agent_name}' ({top.composite_rank_score:.2f}) "
f"vs '{bottom.agent_name}' ({bottom.composite_rank_score:.2f})"
)
if top.generalization_score > 0.7:
insights.append(
f"'{top.agent_name}' shows strong generalization "
f"(std={top.std_score:.3f} across {top.tasks_run} tasks)"
)
for r in self.rankings:
if r.mean_causal_score > 0.6:
insights.append(
f"'{r.agent_name}' demonstrated genuine causal reasoning "
f"(causal_score={r.mean_causal_score:.2f})"
)
strategies = [r.dominant_strategy for r in self.rankings]
if len(set(strategies)) > 1:
best_strategy = self.rankings[0].dominant_strategy
insights.append(
f"Strategy '{best_strategy}' produced the highest composite score."
)
return insights
def render_table(self) -> str:
"""Render ASCII leaderboard table."""
if not self.rankings:
return "No results."
lines = [
f"{'β'*90}",
f" π BENCHMARK LEADERBOARD β {self.benchmark_id}",
f" Tasks: {', '.join(self.tasks_evaluated)} | Agents: {len(self.agents_evaluated)} | Episodes: {self.total_episodes}",
f"{'β'*90}",
f"{'Rank':<5} {'Agent':<16} {'Score':<8} {'Causal':<8} {'Robust':<8} {'Calibr':<8} {'Genrz':<8} {'Steps':<7} {'Strategy'}",
f"{'β'*90}",
]
for i, r in enumerate(self.rankings):
medal = "π₯" if i == 0 else "π₯" if i == 1 else "π₯" if i == 2 else f" #{i+1}"
lines.append(
f"{medal:<5} {r.agent_name:<16} {r.mean_score:<8.3f} "
f"{r.mean_causal_score:<8.3f} {r.mean_robustness_score:<8.3f} "
f"{r.mean_calibration_score:<8.3f} {r.generalization_score:<8.3f} "
f"{r.mean_steps:<7.1f} {r.dominant_strategy}"
)
lines.append(f"{'β'*90}")
lines.append("\nπ Per-Task Breakdown:")
for r in self.rankings:
task_str = " | ".join(f"{t}: {s:.2f}" for t, s in sorted(r.per_task_scores.items()))
lines.append(f" {r.agent_name:<16} {task_str}")
if self._generate_insights():
lines.append("\nπ‘ Insights:")
lines.extend(f" β {i}" for i in self._generate_insights())
return "\n".join(lines)
class BenchmarkRunner:
"""
Automated benchmark runner.
Runs each agent in AGENT_CONFIGS across each task, collecting:
- Final score
- All intelligence metrics (causal, counterfactual, confidence)
- Strategy and failure classification
- Reliability index
Then generates a ranked leaderboard.
"""
def run(
self,
env,
tasks: Optional[List[str]] = None,
agents: Optional[List[str]] = None,
benchmark_id: Optional[str] = None,
) -> LeaderboardReport:
"""Run the full benchmark."""
import uuid
from server.models import RepoAction
from server.strategy_detector import StrategyDetector
from server.failure_classifier import FailureClassifier
from server.advanced_metrics import AdvancedMetricsEngine
from server.causal_probe import CausalProbe
from server.counterfactual_engine import CounterfactualEngine
from server.confidence_calibrator import ConfidenceCalibrator
benchmark_id = benchmark_id or f"bench_{uuid.uuid4().hex[:8]}"
tasks = tasks or ["task1", "task2", "task3"]
agent_configs = self._get_agent_configs()
if agents:
agent_configs = {k: v for k, v in agent_configs.items() if k in agents}
clf = FailureClassifier()
det = StrategyDetector()
adv = AdvancedMetricsEngine()
causal = CausalProbe()
counter = CounterfactualEngine()
calibrator = ConfidenceCalibrator()
start_time = time.time()
all_results: List[BenchmarkResult] = []
for task in tasks:
for agent_name, agent_fn in agent_configs.items():
try:
result = self._run_episode(
env, task, agent_name, agent_fn,
clf, det, adv, causal, counter, calibrator
)
all_results.append(result)
except Exception as e:
# Don't crash the whole benchmark on one failure
all_results.append(BenchmarkResult(
agent_name=agent_name, task=task, variant_id="error",
final_score=0.0, total_steps=0, cumulative_reward=0.0,
duration_seconds=0.0, strategy="ERROR", failure_type="BENCHMARK_ERROR",
reliability_index=0.0, causal_score=0.0, robustness_score=0.0,
calibration_score=0.0, action_sequence=[],
))
total_duration = time.time() - start_time
rankings = self._compute_rankings(all_results, tasks)
return LeaderboardReport(
benchmark_id=benchmark_id,
tasks_evaluated=tasks,
agents_evaluated=list(agent_configs.keys()),
total_episodes=len(all_results),
run_duration_seconds=total_duration,
rankings=rankings,
raw_results=all_results,
)
def _run_episode(
self, env, task, agent_name, agent_fn,
clf, det, adv, causal, counter, calibrator
) -> BenchmarkResult:
from server.models import RepoAction
reset_result = env.reset(task=task)
obs = reset_result.observation
variant_id = reset_result.info.get("variant_id", "unknown")
context = {}
obs_dict = obs.model_dump()
start = time.time()
cumulative_reward = 0.0
files_read, files_written, action_sequence = [], [], []
max_steps = 15
for step_num in range(1, max_steps + 1):
if env.done:
break
action_dict = agent_fn(obs_dict, step_num, context)
action = RepoAction(
action_type=action_dict.get("action_type", "submit"),
path=action_dict.get("path"),
query=action_dict.get("query"),
)
result = env.step(action)
obs = result.observation
obs_dict = obs.model_dump()
cumulative_reward += result.reward
action_sequence.append(action.action_type)
if action.path and action.action_type == "read_file":
files_read.append(action.path)
if action.path and action.action_type == "write_file":
files_written.append(action.path)
if result.done:
break
if not env.done:
r = env.step(RepoAction(action_type="submit"))
cumulative_reward += r.reward
action_sequence.append("submit")
duration = time.time() - start
final_score = env.final_score
traj = env.get_trajectory()
steps = traj.get("steps", []) if traj else []
meta = env.variant.meta if env.variant else {}
# Intelligence metrics
fail_r = clf.classify(
traj.get("episode_id", "") if traj else "", task,
steps, meta, files_read, files_written, final_score
)
strat_r = det.detect(steps, task, meta, files_read, final_score)
adv_r = adv.compute(steps, meta, final_score, files_read, files_written)
causal_r = causal.probe(
traj.get("episode_id", "") if traj else "", task,
steps, meta, files_read, files_written, final_score
)
counter_r = counter.analyze(
traj.get("episode_id", "") if traj else "", task,
steps, meta, files_read, files_written, final_score
)
calib_r = calibrator.calibrate(
traj.get("episode_id", "") if traj else "", task,
steps, final_score,
)
return BenchmarkResult(
agent_name=agent_name,
task=task,
variant_id=variant_id,
final_score=final_score,
total_steps=len(action_sequence),
cumulative_reward=cumulative_reward,
duration_seconds=duration,
strategy=strat_r.strategy,
failure_type=fail_r.primary_failure,
reliability_index=adv_r.reliability_index,
causal_score=causal_r.causal_score,
robustness_score=counter_r.robustness_score,
calibration_score=calib_r.calibration_score,
action_sequence=action_sequence,
)
def _compute_rankings(
self, results: List[BenchmarkResult], tasks: List[str]
) -> List[AgentBenchmarkSummary]:
import math
from collections import Counter
# Group by agent
agent_results: Dict[str, List[BenchmarkResult]] = {}
for r in results:
agent_results.setdefault(r.agent_name, []).append(r)
summaries = []
for agent_name, agent_res in agent_results.items():
scores = [r.final_score for r in agent_res]
mean_score = sum(scores) / len(scores)
if len(scores) > 1:
variance = sum((s - mean_score) ** 2 for s in scores) / len(scores)
std_score = math.sqrt(variance)
else:
std_score = 0.0
generalization_score = max(0.0, 1.0 - std_score)
per_task = {r.task: r.final_score for r in agent_res}
strategies = Counter(r.strategy for r in agent_res)
failures = Counter(r.failure_type for r in agent_res)
mean_steps = sum(r.total_steps for r in agent_res) / len(agent_res)
mean_reliability = sum(r.reliability_index for r in agent_res) / len(agent_res)
mean_causal = sum(r.causal_score for r in agent_res) / len(agent_res)
mean_robustness = sum(r.robustness_score for r in agent_res) / len(agent_res)
mean_calibration = sum(r.calibration_score for r in agent_res) / len(agent_res)
# Composite leaderboard score β weighted across all dimensions
composite = (
mean_score * 0.35 +
mean_causal * 0.20 +
mean_robustness * 0.15 +
mean_calibration * 0.15 +
generalization_score * 0.15
)
best_task = max(per_task, key=per_task.get)
worst_task = min(per_task, key=per_task.get)
summaries.append(AgentBenchmarkSummary(
agent_name=agent_name,
tasks_run=len(agent_res),
mean_score=mean_score,
std_score=std_score,
generalization_score=generalization_score,
mean_steps=mean_steps,
best_task=best_task,
worst_task=worst_task,
mean_reliability=mean_reliability,
mean_causal_score=mean_causal,
mean_robustness_score=mean_robustness,
mean_calibration_score=mean_calibration,
dominant_strategy=strategies.most_common(1)[0][0],
dominant_failure=failures.most_common(1)[0][0],
composite_rank_score=composite,
per_task_scores=per_task,
))
summaries.sort(key=lambda s: -s.composite_rank_score)
return summaries
def _get_agent_configs(self) -> Dict:
"""Reuse built-in strategies from multi_agent.py."""
from server.multi_agent import MultiAgentComparison
return MultiAgentComparison.AGENT_CONFIGS
|