Spaces:
Sleeping
Sleeping
File size: 22,184 Bytes
637f42c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 | #!/usr/bin/env python3
"""
TeamForge Benchmark Runner
==========================
Evaluates any OpenAI-compatible LLM against all TeamForge tasks and produces
a standardized leaderboard-style report β identical protocol every run.
Usage:
python benchmark.py --model llama3-8b-8192
python benchmark.py --model llama3-70b-8192 --model llama3-8b-8192
python benchmark.py --compare # compare all models in results/
Output:
results/<model>/<timestamp>.json β raw episode data
results/leaderboard.json β aggregate leaderboard
results/leaderboard.md β markdown table (paste into README)
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Dict, List, Optional
from openai import OpenAI
from rich.console import Console
from rich.live import Live
from rich.panel import Panel
from rich.progress import BarColumn, Progress, SpinnerColumn, TextColumn, TimeElapsedColumn
from rich.table import Table
from rich import box
from environment import TeamForgeEnv
from models import (
Commit, EditFile, GenerateReview, Observation,
PlanStep, RequestIteration, RunLint, RunTests, SelfReflect,
)
from tasks import ALL_TASK_IDS
console = Console()
# βββ Config ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
GROQ_API_KEY = os.getenv("GROQ_API_KEY", "")
API_BASE_URL = os.getenv("API_BASE_URL", "https://api.groq.com/openai/v1")
RESULTS_DIR = Path("results")
RESULTS_DIR.mkdir(exist_ok=True)
BENCHMARK_VERSION = "1.0.0"
TASK_WEIGHTS = {
"easy_bugfix_chunk_list": 0.20,
"medium_refactor_stats": 0.35,
"hard_lru_cache_performance": 0.45,
}
# βββ System prompt (CoT + structured) ββββββββββββββββββββββββββββββββββββββββ
SYSTEM_PROMPT = """\
You are TeamForge-Agent, an autonomous AI software engineer.
You work through tasks in strict phases:
PLAN β CODE β TEST β LINT β REVIEW β REFLECT β COMMIT
βββ REASONING PROTOCOL βββ
Before each action, reason silently:
OBSERVE: What does the current state tell me?
DIAGNOSE: What is the root cause / what needs doing?
DECIDE: Which action type moves me closer to done?
EXECUTE: Return the action JSON.
βββ ACTION RULES βββ
β’ NEVER modify files whose path contains "test"
β’ Always emit β₯2 plan_step actions before any edit_file
β’ After every edit_file, run_tests before committing
β’ generate_review must mention specific code details (not generic praise)
β’ Commit message must follow Conventional Commits: fix/feat/refactor/perf(scope): desc
β’ Return ONLY valid JSON β no markdown, no explanation outside the JSON
βββ AVAILABLE ACTIONS βββ
{"type":"plan_step", "step_number":1, "description":"...", "estimated_effort":"low|medium|high"}
{"type":"edit_file", "file_path":"...", "content":"<full file>", "reason":"..."}
{"type":"run_tests", "timeout_seconds":30}
{"type":"run_lint", "fix":false}
{"type":"generate_review", "focus_areas":["correctness","performance"], "review_text":"..."}
{"type":"commit", "message":"fix(scope): ..."}
{"type":"self_reflect", "what_went_well":"...", "what_to_improve":"..."}
{"type":"request_iteration","reason":"...", "target_issues":["..."]}
"""
# βββ Agent βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class BenchmarkAgent:
def __init__(self, client: OpenAI, model: str):
self.client = client
self.model = model
self.history: List[Dict] = []
self._consecutive_failures = 0
def reset(self):
self.history = []
self._consecutive_failures = 0
def act(self, obs: Observation) -> Optional[Any]:
self.history.append({"role": "user", "content": self._obs_to_prompt(obs)})
for attempt in range(3):
try:
resp = self.client.chat.completions.create(
model=self.model,
messages=[{"role":"system","content":SYSTEM_PROMPT}, *self.history[-12:]],
temperature=0.15,
max_tokens=1800,
)
raw = resp.choices[0].message.content.strip()
self.history.append({"role":"assistant","content":raw})
action = self._parse(raw)
if action:
self._consecutive_failures = 0
return action
except Exception as e:
time.sleep(1.5 ** attempt)
self._consecutive_failures += 1
return None
def _obs_to_prompt(self, obs: Observation) -> str:
parts = [
f"TASK: {obs.task_id} | STEP {obs.step_number}/{obs.max_steps} | PHASE: {obs.phase.value}",
f"CUMULATIVE_REWARD: {obs.cumulative_reward:.3f}",
f"\n## TASK\n{obs.task_description[:500]}",
]
if obs.last_action_type:
parts.append(f"\n## LAST ACTION: {obs.last_action_type} β {obs.last_action_status.value}")
parts.append(f"```\n{obs.last_action_output[:600]}\n```")
if obs.test_results:
t = obs.test_results
parts.append(f"\n## TESTS: {t.passed} passed / {t.failed} failed / {t.errors} errors")
if t.failed or t.errors:
parts.append(f"```\n{t.output[-600:]}\n```")
if obs.lint_results and obs.lint_results.violations:
parts.append(f"\n## LINT: {obs.lint_results.violations} violations\n```\n{obs.lint_results.output[:400]}\n```")
parts.append("\n## REPO FILES")
for f in obs.repo_files[:8]:
if f.size_bytes < 5000:
parts.append(f"\n### {f.path}\n```python\n{f.content[:1000]}\n```")
if obs.plan:
parts.append(f"\n## PLAN ({len(obs.plan)} steps)")
for s in obs.plan:
parts.append(f" {s.step_number}. [{s.estimated_effort}] {s.description}")
parts.append("\n## NEXT ACTION (JSON only):")
return "\n".join(parts)
def _parse(self, text: str) -> Optional[Any]:
text = text.strip().lstrip("```json").lstrip("```").rstrip("```").strip()
from models import PlanStep, EditFile, RunTests, RunLint, GenerateReview, Commit, SelfReflect, RequestIteration
dispatch = {
"plan_step": PlanStep, "edit_file": EditFile, "run_tests": RunTests,
"run_lint": RunLint, "generate_review": GenerateReview, "commit": Commit,
"self_reflect": SelfReflect, "request_iteration": RequestIteration,
}
try:
data = json.loads(text)
cls = dispatch.get(data.get("type",""))
return cls(**data) if cls else None
except Exception:
# Try to extract JSON from response
import re
m = re.search(r'\{.*\}', text, re.DOTALL)
if m:
try:
data = json.loads(m.group())
cls = dispatch.get(data.get("type",""))
return cls(**data) if cls else None
except Exception:
pass
return None
# βββ Episode runner βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def run_episode(
env: TeamForgeEnv,
agent: BenchmarkAgent,
task_id: str,
) -> Dict[str, Any]:
agent.reset()
obs = env.reset(task_id)
steps_log = []
start_time = time.perf_counter()
while not obs.done:
t0 = time.perf_counter()
action = agent.act(obs)
latency = time.perf_counter() - t0
if action is None:
break
obs = env.step(action)
steps_log.append({
"step": obs.step_number,
"action": obs.last_action_type,
"status": obs.last_action_status.value,
"reward": obs.reward,
"cum_reward": obs.cumulative_reward,
"latency_s": round(latency, 3),
"tests_passed": obs.test_results.passed if obs.test_results else 0,
"tests_failed": obs.test_results.failed if obs.test_results else 0,
"lint_violations": obs.lint_results.violations if obs.lint_results else None,
})
elapsed = time.perf_counter() - start_time
result = env.grade()
return {
"task_id": task_id,
"model": agent.model,
"timestamp": datetime.now(timezone.utc).isoformat(),
"elapsed_s": round(elapsed, 2),
"total_steps": obs.step_number,
"done": obs.done,
"steps_log": steps_log,
"scores": {
"test_pass_rate": result.test_pass_rate,
"lint_score": result.lint_score,
"efficiency_score": result.efficiency_score,
"review_quality": result.review_quality,
"reflection_quality": result.reflection_quality,
"final_score": result.final_score,
"passed": result.passed,
},
}
# βββ Leaderboard βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_teamforge_score(results_by_task: Dict[str, Dict]) -> float:
"""Weighted composite score across all tasks (0.0β1.0)."""
total = sum(
TASK_WEIGHTS.get(tid, 0) * r["scores"]["final_score"]
for tid, r in results_by_task.items()
)
return round(total, 4)
def build_leaderboard() -> List[Dict]:
"""Scan results/ and aggregate one row per model."""
rows = []
for model_dir in RESULTS_DIR.iterdir():
if not model_dir.is_dir():
continue
model_runs: Dict[str, List] = {}
for f in model_dir.glob("*.json"):
try:
data = json.loads(f.read_text())
tid = data["task_id"]
model_runs.setdefault(tid, []).append(data)
except Exception:
pass
if not model_runs:
continue
# Best run per task
best = {}
for tid, runs in model_runs.items():
best[tid] = max(runs, key=lambda r: r["scores"]["final_score"])
ts = compute_teamforge_score(best)
rows.append({
"model": model_dir.name,
"teamforge_score": ts,
"tasks": {
tid: r["scores"] for tid, r in best.items()
},
"avg_steps": round(
sum(r["total_steps"] for r in best.values()) / max(1, len(best)), 1
),
})
rows.sort(key=lambda r: r["teamforge_score"], reverse=True)
return rows
def save_leaderboard(rows: List[Dict]) -> None:
(RESULTS_DIR / "leaderboard.json").write_text(json.dumps(rows, indent=2))
# Markdown table
lines = [
"## TeamForge Leaderboard\n",
"| Rank | Model | TeamForge Score | Easy | Medium | Hard | Avg Steps |",
"|------|-------|:--------------:|:----:|:------:|:----:|:---------:|",
]
for i, r in enumerate(rows, 1):
easy = r["tasks"].get("easy_bugfix_chunk_list", {}).get("final_score", 0)
medium = r["tasks"].get("medium_refactor_stats", {}).get("final_score", 0)
hard = r["tasks"].get("hard_lru_cache_performance",{}).get("final_score", 0)
lines.append(
f"| {i} | `{r['model']}` | **{r['teamforge_score']:.4f}** "
f"| {easy:.3f} | {medium:.3f} | {hard:.3f} | {r['avg_steps']} |"
)
(RESULTS_DIR / "leaderboard.md").write_text("\n".join(lines))
# βββ Rich display βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def print_episode_live(task_id: str, model: str, episode: Dict) -> None:
scores = episode["scores"]
log = episode["steps_log"]
# Step table
step_table = Table(box=box.SIMPLE, show_header=True, header_style="bold cyan")
step_table.add_column("Step", justify="right", width=5)
step_table.add_column("Action", width=20)
step_table.add_column("Status", width=9)
step_table.add_column("Reward", justify="right", width=8)
step_table.add_column("Cum Reward", justify="right", width=10)
step_table.add_column("Tests", justify="right", width=9)
for s in log:
status_style = "green" if s["status"] == "success" else "red"
reward_style = "green" if s["reward"] > 0 else ("red" if s["reward"] < -0.05 else "dim")
step_table.add_row(
str(s["step"]),
s["action"],
f"[{status_style}]{s['status']}[/{status_style}]",
f"[{reward_style}]{s['reward']:+.4f}[/{reward_style}]",
f"{s['cum_reward']:+.4f}",
f"{s['tests_passed']}p / {s['tests_failed']}f",
)
# Score panel
score_table = Table(box=box.SIMPLE, show_header=False)
score_table.add_column("Metric", style="bold")
score_table.add_column("Score", justify="right")
score_table.add_column("Bar", width=20)
def bar(v: float) -> str:
filled = int(v * 20)
color = "green" if v >= 0.8 else ("yellow" if v >= 0.5 else "red")
return f"[{color}]{'β' * filled}{'β' * (20 - filled)}[/{color}]"
rows_s = [
("test_pass_rate", scores["test_pass_rate"], "Tests Passed"),
("lint_score", scores["lint_score"], "Lint Clean"),
("efficiency", scores["efficiency_score"], "Efficiency"),
("review_quality", scores["review_quality"], "Review Quality"),
("reflection", scores["reflection_quality"],"Reflection"),
]
for _, v, label in rows_s:
score_table.add_row(label, f"{v:.4f}", bar(v))
passed_str = "[bold green]β PASSED[/bold green]" if scores["passed"] else "[bold red]β FAILED[/bold red]"
console.print(Panel(
step_table,
title=f"[bold]{task_id}[/bold] [{model}]",
subtitle=f"Final: [bold yellow]{scores['final_score']:.4f}[/bold yellow] {passed_str} "
f"Steps: {episode['total_steps']} Time: {episode['elapsed_s']:.1f}s",
border_style="blue",
))
console.print(score_table)
def print_leaderboard(rows: List[Dict]) -> None:
table = Table(
title="[bold]TeamForge Leaderboard[/bold]",
box=box.DOUBLE_EDGE,
show_header=True,
header_style="bold magenta",
border_style="magenta",
)
table.add_column("Rank", justify="center", width=5)
table.add_column("Model", width=28)
table.add_column("TeamForge Score", justify="center", width=16)
table.add_column("Easy (20%)", justify="center", width=11)
table.add_column("Medium (35%)", justify="center", width=12)
table.add_column("Hard (45%)", justify="center", width=11)
table.add_column("Avg Steps", justify="center", width=10)
medals = ["π₯","π₯","π₯"]
for i, r in enumerate(rows):
medal = medals[i] if i < 3 else f"#{i+1}"
easy = r["tasks"].get("easy_bugfix_chunk_list", {}).get("final_score", 0)
medium = r["tasks"].get("medium_refactor_stats", {}).get("final_score", 0)
hard = r["tasks"].get("hard_lru_cache_performance",{}).get("final_score", 0)
ts = r["teamforge_score"]
ts_color = "green" if ts >= 0.80 else ("yellow" if ts >= 0.55 else "red")
table.add_row(
medal,
f"[bold]{r['model']}[/bold]",
f"[{ts_color}]{ts:.4f}[/{ts_color}]",
f"{easy:.3f}",
f"{medium:.3f}",
f"{hard:.3f}",
str(r["avg_steps"]),
)
console.print(table)
# βββ Main βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser(description="TeamForge Benchmark Runner")
parser.add_argument("--model", action="append", dest="models",
default=None, help="Model name (repeat for multiple)")
parser.add_argument("--task", choices=ALL_TASK_IDS + ["all"], default="all")
parser.add_argument("--compare", action="store_true",
help="Only rebuild + print leaderboard from saved results")
args = parser.parse_args()
if args.compare:
rows = build_leaderboard()
save_leaderboard(rows)
print_leaderboard(rows)
return
models = args.models or ["llama3-8b-8192"]
task_ids = ALL_TASK_IDS if args.task == "all" else [args.task]
console.rule("[bold blue]TeamForge Benchmark[/bold blue]")
console.print(f"[dim]Version {BENCHMARK_VERSION} | {len(models)} model(s) Γ {len(task_ids)} task(s)[/dim]\n")
client = OpenAI(api_key=GROQ_API_KEY, base_url=API_BASE_URL)
for model in models:
model_dir = RESULTS_DIR / model.replace("/","_")
model_dir.mkdir(exist_ok=True)
agent = BenchmarkAgent(client, model)
env = TeamForgeEnv()
console.print(f"\n[bold cyan]βΆ Model: {model}[/bold cyan]")
for task_id in task_ids:
with Progress(
SpinnerColumn(),
TextColumn(f"[bold]{task_id}[/bold]"),
BarColumn(),
TimeElapsedColumn(),
console=console,
transient=True,
) as progress:
progress.add_task("running", total=None)
episode = run_episode(env, agent, task_id)
# Save raw result
ts_str = datetime.now().strftime("%Y%m%d_%H%M%S")
out_file = model_dir / f"{task_id}_{ts_str}.json"
out_file.write_text(json.dumps(episode, indent=2))
print_episode_live(task_id, model, episode)
env._sandbox.teardown()
# Rebuild leaderboard
rows = build_leaderboard()
save_leaderboard(rows)
console.print()
console.rule("[bold magenta]Leaderboard[/bold magenta]")
print_leaderboard(rows)
console.print(f"\n[dim]Full results β {RESULTS_DIR}/leaderboard.json[/dim]")
console.print(f"[dim]Markdown table β {RESULTS_DIR}/leaderboard.md[/dim]")
if __name__ == "__main__":
main()
# βββ Failure Mode Analyzer ββββββββββββββββββββββββββββββββββββββββββββββββββββ
def analyze_failures(episodes: list[dict]) -> dict:
"""
Analyze failure modes across episodes.
Returns a dict of {failure_type: count}.
"""
modes = {
"test_file_modified": 0,
"max_steps_exceeded": 0,
"zero_tests_passed": 0,
"lint_never_run": 0,
"no_review_written": 0,
"no_plan_written": 0,
"immediate_edit_no_plan": 0,
}
for ep in episodes:
log = ep.get("steps_log", [])
scores = ep.get("scores", {})
actions = [s["action"] for s in log]
if not ep.get("done"):
modes["max_steps_exceeded"] += 1
if scores.get("test_pass_rate", 1) == 0:
modes["zero_tests_passed"] += 1
if "run_lint" not in actions:
modes["lint_never_run"] += 1
if "generate_review" not in actions:
modes["no_review_written"] += 1
if "plan_step" not in actions:
modes["no_plan_written"] += 1
# Check if first code action was edit without prior plan
first_edit = next((i for i, a in enumerate(actions) if a == "edit_file"), 999)
first_plan = next((i for i, a in enumerate(actions) if a == "plan_step"), 999)
if first_edit < first_plan:
modes["immediate_edit_no_plan"] += 1
# Test file modification = reward < -0.25 on edit_file step
for s in log:
if s["action"] == "edit_file" and s["reward"] < -0.25:
modes["test_file_modified"] += 1
break
return {k: v for k, v in modes.items() if v > 0}
def print_failure_analysis(episodes: list[dict], model: str) -> None:
modes = analyze_failures(episodes)
if not modes:
console.print(f" [green]No failure modes detected for {model}[/green]")
return
table = Table(
title=f"[bold red]Failure Mode Analysis β {model}[/bold red]",
box=box.SIMPLE,
show_header=True,
header_style="bold red",
)
table.add_column("Failure Mode", width=30)
table.add_column("Count", justify="right", width=7)
table.add_column("Implication", width=40)
implications = {
"test_file_modified": "Agent tried to cheat β severe penalty applied",
"max_steps_exceeded": "Agent ran out of steps without completing",
"zero_tests_passed": "Implementation entirely wrong / syntax error",
"lint_never_run": "Agent skipped code quality check",
"no_review_written": "Agent skipped review phase",
"no_plan_written": "Agent dove straight into coding without planning",
"immediate_edit_no_plan": "First edit came before first plan step",
}
for mode, count in sorted(modes.items(), key=lambda x: -x[1]):
table.add_row(
mode.replace("_", " "),
str(count),
implications.get(mode, ""),
)
console.print(table)
|