Spaces:
Sleeping
Sleeping
File size: 11,417 Bytes
637f42c a28d168 637f42c e317eba 4f893da e317eba 637f42c 8842df9 637f42c e317eba 637f42c 8842df9 637f42c efa2d2a 637f42c 8842df9 637f42c efa2d2a 637f42c e317eba 637f42c 8842df9 637f42c efa2d2a 637f42c e317eba 637f42c e317eba 637f42c e317eba 637f42c e317eba 637f42c e317eba 637f42c 94d08ee 4f893da | 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 | """
TeamForge Grader
Deterministic scoring of a completed episode.
Final score formula (0.0β1.0):
score = 0.40 * test_pass_rate
+ 0.25 * lint_score
+ 0.20 * efficiency_score
+ 0.10 * review_quality
+ 0.05 * reflection_quality
Anti-exploit measures:
- Test tampering detection (trivial pass/assert True bodies)
- Implementation existence check (no empty stubs)
- Minimum review length requirement
- Lint-fix-only exploit detection
"""
from __future__ import annotations
import ast
import math
import re
import subprocess
import sys
from pathlib import Path
from typing import List, Optional
from models import EpisodeResult, ReviewArtifact, ReflectionArtifact
# βββββββββββββββββββββββββββββββββββββββββββββ
# SCORING CONFIG
# βββββββββββββββββββββββββββββββββββββββββββββ
_SCORE_MIN = 0.10 # deep interior to avoid all boundary issues
_SCORE_MAX = 0.90 # deep interior to avoid all boundary issues
def _clamp(score: float) -> float:
"""Ensure score is strictly within the open interval (0, 1)."""
return round(max(_SCORE_MIN, min(_SCORE_MAX, score)), 4)
# βββββββββββββββββββββββββββββββββββββββββββββ
# ANTI-EXPLOIT GUARDS
# βββββββββββββββββββββββββββββββββββββββββββββ
def _detect_test_tampering(repo_path: str) -> bool:
"""
Detect trivially-passing test rewrites.
Flags: single-statement test with `pass`, `return`, or `assert True`.
"""
for test_file in Path(repo_path).rglob("test_*.py"):
try:
src = test_file.read_text()
tree = ast.parse(src)
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef) and node.name.startswith("test_"):
body = node.body
if len(body) == 1:
stmt = body[0]
if isinstance(stmt, (ast.Pass, ast.Return)):
return True
if isinstance(stmt, ast.Assert):
val = stmt.test
if isinstance(val, ast.Constant) and val.value is True:
return True
except Exception:
pass
return False
def _implementation_exists(repo_path: str) -> bool:
"""Verify at least some non-test code was written (>5 non-blank lines)."""
total = 0
for f in Path(repo_path).rglob("*.py"):
if "test" in f.name or "__pycache__" in str(f):
continue
try:
lines = [l for l in f.read_text().splitlines() if l.strip()]
total += len(lines)
except Exception:
pass
return total >= 5
# βββββββββββββββββββββββββββββββββββββββββββββ
# SUB-SCORERS
# βββββββββββββββββββββββββββββββββββββββββββββ
def score_tests(repo_path: str, timeout: int = 60) -> tuple[float, str]:
"""Run pytest and return (pass_rate 0-1, output)."""
result = subprocess.run(
[sys.executable, "-m", "pytest", "--tb=short", "-q", "--no-header"],
cwd=repo_path,
capture_output=True,
text=True,
timeout=timeout,
)
output = result.stdout + result.stderr
passed = failed = errors = 0
m_p = re.search(r"(\d+) passed", output)
m_f = re.search(r"(\d+) failed", output)
m_e = re.search(r"(\d+) error", output)
if m_p: passed = int(m_p.group(1))
if m_f: failed = int(m_f.group(1))
if m_e: errors = int(m_e.group(1))
total = passed + failed + errors
if total == 0:
return _SCORE_MIN, output
pass_rate = passed / total
return _clamp(pass_rate), output
def score_lint(repo_path: str) -> tuple[float, str]:
"""Run ruff and return (lint_score 0-1, output)."""
result = subprocess.run(
[sys.executable, "-m", "ruff", "check", "."],
cwd=repo_path,
capture_output=True,
text=True,
)
output = result.stdout + result.stderr
if result.returncode == 0:
return _SCORE_MAX, "No lint violations."
violations = len([
ln for ln in output.splitlines()
if re.match(r".+:\d+:\d+:", ln)
])
# Strictly interior [SCORE_MIN, SCORE_MAX]
raw_score = 1.0 - violations * 0.07
return _clamp(raw_score), output
def score_review_quality(
reviews: List[ReviewArtifact],
required_keywords: List[str],
) -> float:
"""Keyword-based review quality with minimum length requirement."""
if not reviews:
return _SCORE_MIN
combined = " ".join(r.text.lower() for r in reviews)
# Anti-exploit: minimum meaningful length
if len(combined.strip()) < 40:
return _SCORE_MIN + 0.05
# Keyword coverage
if not required_keywords:
kw_score = 0.5
else:
found = sum(1 for kw in required_keywords if kw.lower() in combined)
kw_score = found / len(required_keywords)
# Length bonus (up to 0.2 for thorough reviews)
avg_len = sum(len(r.text) for r in reviews) / len(reviews)
length_bonus = min(0.2, avg_len / 600)
# Specificity bonus: mentions actual code identifiers
code_words = re.findall(r'\b[a-z_]{3,}\(\)', combined)
specificity = min(0.1, len(set(code_words)) * 0.025)
return _clamp(kw_score * 0.7 + length_bonus + specificity)
def score_reflection_quality(reflections: List[ReflectionArtifact]) -> float:
"""Score reflections on depth and actionability."""
if not reflections:
return _SCORE_MIN
total = 0.0
for ref in reflections:
depth = 0.0
if len(ref.what_went_well.strip()) > 40:
depth += 0.5
if len(ref.what_to_improve.strip()) > 40:
depth += 0.5
# Bonus if adjusted_plan provided
if ref.adjusted_plan and len(ref.adjusted_plan.strip()) > 20:
depth = min(0.9, depth + 0.2)
total += depth
# Strictly (0, 1) - Safer interior
return _clamp(total / max(1, len(reflections)))
def score_efficiency(total_steps: int, max_steps: int) -> float:
"""Reward solving in fewer steps with smooth decay curve."""
ratio = total_steps / max_steps
# Strictly (0, 1) - Safer interior
return _clamp(math.exp(-2.0 * max(0, ratio - 0.25)))
# βββββββββββββββββββββββββββββββββββββββββββββ
# MAIN GRADER
# βββββββββββββββββββββββββββββββββββββββββββββ
def grade_episode(
repo_path: str,
task_id: str,
total_steps: int,
max_steps: int,
reviews: List[ReviewArtifact],
reflections: List[ReflectionArtifact],
required_keywords: Optional[List[str]] = None,
) -> EpisodeResult:
"""
Run full deterministic grading pipeline.
Returns EpisodeResult with final_score in [0.0, 1.0].
"""
log: List[str] = []
required_keywords = required_keywords or []
# ββ Anti-exploit checks ββ
if _detect_test_tampering(repo_path):
log.append("[GRADER] β TEST TAMPERING DETECTED β score zeroed")
return EpisodeResult(
task_id=task_id, total_steps=total_steps,
final_score=_SCORE_MIN, passed=False,
log=log + ["Test files were trivially rewritten to force passes."],
)
if not _implementation_exists(repo_path):
log.append("[GRADER] β NO IMPLEMENTATION FOUND β score zeroed")
return EpisodeResult(
task_id=task_id, total_steps=total_steps,
final_score=_SCORE_MIN, passed=False,
log=log + ["No non-test code was written."],
)
# ββ Tests ββ
test_pass_rate, test_output = score_tests(repo_path)
log.append(f"[GRADER] test_pass_rate={test_pass_rate:.3f}")
log.append(f"[GRADER] test_output=\n{test_output[:800]}")
# ββ Lint ββ
lint_score, lint_output = score_lint(repo_path)
log.append(f"[GRADER] lint_score={lint_score:.3f}")
# ββ Efficiency ββ
efficiency = score_efficiency(total_steps, max_steps)
log.append(f"[GRADER] efficiency={efficiency:.3f} ({total_steps}/{max_steps} steps)")
# ββ Review quality ββ
review_q = score_review_quality(reviews, required_keywords)
log.append(f"[GRADER] review_quality={review_q:.3f}")
# ββ Reflection quality ββ
reflect_q = score_reflection_quality(reflections)
log.append(f"[GRADER] reflection_quality={reflect_q:.3f}")
# ββ Final weighted score ββ
final = (
0.40 * test_pass_rate
+ 0.25 * lint_score
+ 0.20 * efficiency
+ 0.10 * review_q
+ 0.05 * reflect_q
)
# Strictly (0, 1) interior range to satisfy Phase 2 validator
final = _clamp(final)
log.append(f"[GRADER] FINAL_SCORE={final:.4f}")
return EpisodeResult(
task_id=task_id,
total_steps=total_steps,
test_pass_rate=test_pass_rate,
lint_score=lint_score,
efficiency_score=efficiency,
review_quality=review_q,
reflection_quality=reflect_q,
final_score=final,
passed=test_pass_rate >= 0.9 and lint_score >= 0.7,
log=log,
)
def grade_task(repo_path: str, **kwargs) -> float:
"""
OpenEnv standard grader bridge β entry point from YAML.
Returns ONLY a float strictly between 0 and 1.
"""
import json
import os
from typing import List
from pydantic import TypeAdapter
metadata_path = os.path.join(repo_path, "grading_metadata.json")
# Default fallback values for out-of-band grading
task_id = "unknown"
total_steps = 1
max_steps = 20
reviews = []
reflections = []
required_keywords = []
if os.path.exists(metadata_path):
try:
with open(metadata_path, "r") as f:
meta = json.load(f)
task_id = meta.get("task_id", task_id)
total_steps = meta.get("total_steps", total_steps)
max_steps = meta.get("max_steps", max_steps)
# Use TypeAdapter for robust Pydantic deserialization
from models import ReviewArtifact, ReflectionArtifact
reviews = TypeAdapter(List[ReviewArtifact]).validate_python(meta.get("reviews", []))
reflections = TypeAdapter(List[ReflectionArtifact]).validate_python(meta.get("reflections", []))
required_keywords = meta.get("required_keywords", [])
except Exception:
pass
try:
result = grade_episode(
repo_path=repo_path,
task_id=task_id,
total_steps=total_steps,
max_steps=max_steps,
reviews=reviews,
reflections=reflections,
required_keywords=required_keywords,
)
return float(result.final_score)
except Exception:
return _SCORE_MIN
|