File size: 16,802 Bytes
ed1b365 | 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 | """
Benchmark Runner - loads test prompts, runs/loads responses, scores them,
and produces detailed evaluation reports.
Supports:
- Loading prompts from JSON files in evaluation/prompts/
- Pre-generated response files (JSON mapping prompt -> response)
- Scoring via ReasoningMetrics
- Per-category and overall reports
- Baseline vs trained model comparison
- CLI interface
"""
from __future__ import annotations
import argparse
import json
import os
import sys
from datetime import datetime
from pathlib import Path
from typing import Any, Dict, List, Optional
# Allow running from project root or from evaluation/
_THIS_DIR = Path(__file__).resolve().parent
_PROJECT_ROOT = _THIS_DIR.parent
if str(_PROJECT_ROOT) not in sys.path:
sys.path.insert(0, str(_PROJECT_ROOT))
from evaluation.reasoning_metrics import ReasoningMetrics
# ---------------------------------------------------------------------------
# Benchmark Runner
# ---------------------------------------------------------------------------
class BenchmarkRunner:
"""Load prompts, score responses, produce reports."""
def __init__(
self,
prompts_dir: Optional[str] = None,
metrics: Optional[ReasoningMetrics] = None,
):
self.prompts_dir = Path(prompts_dir) if prompts_dir else _THIS_DIR / "prompts"
self.metrics = metrics or ReasoningMetrics()
self._prompts: Dict[str, List[str]] = {}
self._counterexamples: List[Dict[str, str]] = []
# -- loading -----------------------------------------------------------
def load_prompts(self, filename: str = "reasoning_tests.json") -> Dict[str, List[str]]:
"""Load categorised prompts from a JSON file.
Expected format: {"category": ["prompt1", "prompt2", ...], ...}
"""
path = self.prompts_dir / filename
if not path.exists():
raise FileNotFoundError(f"Prompt file not found: {path}")
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
self._prompts = data
return data
def load_counterexamples(self, filename: str = "counterexample_tests.json") -> List[Dict[str, str]]:
"""Load counterexample test prompts."""
path = self.prompts_dir / filename
if not path.exists():
raise FileNotFoundError(f"Counterexample file not found: {path}")
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
self._counterexamples = data
return data
def load_responses(self, filepath: str) -> Dict[str, str]:
"""Load pre-generated responses from a JSON file.
Expected format: {"prompt_text": "response_text", ...}
"""
with open(filepath, "r", encoding="utf-8") as f:
return json.load(f)
# -- scoring -----------------------------------------------------------
def score_responses(
self,
responses: Dict[str, str],
) -> Dict[str, Any]:
"""Score all responses and organise results by category.
Args:
responses: mapping of prompt text -> response text
Returns:
Dict with per-prompt scores, per-category averages, and overall.
"""
if not self._prompts:
self.load_prompts()
results: Dict[str, Any] = {
"timestamp": datetime.utcnow().isoformat(),
"total_prompts": 0,
"scored_prompts": 0,
"missing_responses": 0,
"categories": {},
"all_scores": [],
}
for category, prompts in self._prompts.items():
cat_scores: List[Dict[str, Any]] = []
for prompt in prompts:
results["total_prompts"] += 1
response = responses.get(prompt)
if response is None:
results["missing_responses"] += 1
continue
scores = self.metrics.score_reasoning(response)
results["scored_prompts"] += 1
entry = {"prompt": prompt, "scores": scores}
cat_scores.append(entry)
results["all_scores"].append(entry)
# Category averages
if cat_scores:
avg = self._average_scores([e["scores"] for e in cat_scores])
else:
avg = {}
results["categories"][category] = {
"prompts_scored": len(cat_scores),
"average_scores": avg,
"details": cat_scores,
}
# Overall averages
if results["all_scores"]:
results["overall"] = self._average_scores(
[e["scores"] for e in results["all_scores"]]
)
else:
results["overall"] = {}
return results
def score_counterexamples(
self,
responses: Dict[str, str],
) -> Dict[str, Any]:
"""Score counterexample responses (should identify wrong reasoning)."""
if not self._counterexamples:
self.load_counterexamples()
results = []
refutations = 0
total = 0
refutation_markers = [
"not true", "incorrect", "misconception", "actually",
"contrary", "doesn't", "does not", "false", "myth",
"wrong", "mistake", "no,", "in fact", "however",
"this is a common", "oversimplification", "nuanced",
"not necessarily", "depends on", "more complex",
]
for item in self._counterexamples:
prompt = item["prompt"]
expected = item.get("expected", "refutation")
response = responses.get(prompt, "")
total += 1
if not response:
results.append({
"prompt": prompt,
"expected": expected,
"responded": False,
"contains_refutation": False,
})
continue
resp_lower = response.lower()
found_refutation = any(m in resp_lower for m in refutation_markers)
if found_refutation and expected == "refutation":
refutations += 1
scores = self.metrics.score_reasoning(response)
results.append({
"prompt": prompt,
"expected": expected,
"responded": True,
"contains_refutation": found_refutation,
"scores": scores,
})
return {
"total": total,
"refutation_rate": round(refutations / max(total, 1), 4),
"details": results,
}
# -- comparison --------------------------------------------------------
def compare_models(
self,
baseline_responses: Dict[str, str],
trained_responses: Dict[str, str],
) -> Dict[str, Any]:
"""Compare baseline vs trained model responses."""
baseline_results = self.score_responses(baseline_responses)
trained_results = self.score_responses(trained_responses)
comparison: Dict[str, Any] = {
"timestamp": datetime.utcnow().isoformat(),
"baseline_overall": baseline_results.get("overall", {}),
"trained_overall": trained_results.get("overall", {}),
"category_comparison": {},
"improvements": {},
"regressions": {},
}
# Per-category delta
for cat in baseline_results["categories"]:
b_avg = baseline_results["categories"][cat]["average_scores"]
t_avg = trained_results["categories"].get(cat, {}).get("average_scores", {})
delta = {}
for k in b_avg:
if k in t_avg and isinstance(b_avg[k], (int, float)):
delta[k] = round(t_avg[k] - b_avg[k], 4)
comparison["category_comparison"][cat] = {
"baseline": b_avg,
"trained": t_avg,
"delta": delta,
}
# Overall delta
b_ov = comparison["baseline_overall"]
t_ov = comparison["trained_overall"]
for k in b_ov:
if k in t_ov and isinstance(b_ov[k], (int, float)):
d = round(t_ov[k] - b_ov[k], 4)
if d > 0.01:
comparison["improvements"][k] = d
elif d < -0.01:
comparison["regressions"][k] = d
return comparison
# -- report ------------------------------------------------------------
def format_report(self, results: Dict[str, Any]) -> str:
"""Format evaluation results as a readable text report."""
lines: List[str] = []
lines.append("=" * 70)
lines.append(" CODETTE BENCHMARK EVALUATION REPORT")
lines.append("=" * 70)
lines.append(f" Timestamp: {results.get('timestamp', 'N/A')}")
lines.append(f" Prompts: {results.get('scored_prompts', 0)} scored / "
f"{results.get('total_prompts', 0)} total")
if results.get("missing_responses"):
lines.append(f" Missing: {results['missing_responses']} responses not found")
lines.append("")
# Overall
overall = results.get("overall", {})
if overall:
lines.append("-" * 70)
lines.append(" OVERALL SCORES")
lines.append("-" * 70)
for k, v in sorted(overall.items()):
if isinstance(v, float):
bar = self._bar(v)
lines.append(f" {k:<22s} {v:.4f} {bar}")
lines.append("")
# Per-category
for cat, data in results.get("categories", {}).items():
avg = data.get("average_scores", {})
if not avg:
continue
lines.append("-" * 70)
lines.append(f" CATEGORY: {cat.upper()}")
lines.append(f" Prompts scored: {data.get('prompts_scored', 0)}")
lines.append("-" * 70)
for k, v in sorted(avg.items()):
if isinstance(v, float):
bar = self._bar(v)
lines.append(f" {k:<22s} {v:.4f} {bar}")
lines.append("")
lines.append("=" * 70)
return "\n".join(lines)
def format_comparison_report(self, comparison: Dict[str, Any]) -> str:
"""Format a comparison report between baseline and trained model."""
lines: List[str] = []
lines.append("=" * 70)
lines.append(" MODEL COMPARISON REPORT")
lines.append("=" * 70)
lines.append(f" Timestamp: {comparison.get('timestamp', 'N/A')}")
lines.append("")
# Overall
lines.append("-" * 70)
lines.append(" OVERALL SCORES (baseline -> trained [delta])")
lines.append("-" * 70)
b = comparison.get("baseline_overall", {})
t = comparison.get("trained_overall", {})
for k in sorted(set(list(b.keys()) + list(t.keys()))):
bv = b.get(k, 0)
tv = t.get(k, 0)
if not isinstance(bv, (int, float)):
continue
d = tv - bv
sign = "+" if d >= 0 else ""
lines.append(f" {k:<22s} {bv:.4f} -> {tv:.4f} [{sign}{d:.4f}]")
# Improvements / regressions
imp = comparison.get("improvements", {})
reg = comparison.get("regressions", {})
if imp:
lines.append("")
lines.append(" IMPROVEMENTS:")
for k, v in sorted(imp.items(), key=lambda x: -x[1]):
lines.append(f" + {k}: +{v:.4f}")
if reg:
lines.append("")
lines.append(" REGRESSIONS:")
for k, v in sorted(reg.items(), key=lambda x: x[1]):
lines.append(f" - {k}: {v:.4f}")
# Per-category
lines.append("")
for cat, data in comparison.get("category_comparison", {}).items():
delta = data.get("delta", {})
if not delta:
continue
overall_d = delta.get("overall", 0)
sign = "+" if overall_d >= 0 else ""
lines.append(f" {cat:<18s} overall delta: {sign}{overall_d:.4f}")
lines.append("")
lines.append("=" * 70)
return "\n".join(lines)
# -- helpers -----------------------------------------------------------
@staticmethod
def _average_scores(score_list: List[Dict[str, float]]) -> Dict[str, float]:
"""Average numeric values across a list of score dicts."""
if not score_list:
return {}
totals: Dict[str, float] = {}
counts: Dict[str, int] = {}
for s in score_list:
for k, v in s.items():
if isinstance(v, (int, float)):
totals[k] = totals.get(k, 0.0) + v
counts[k] = counts.get(k, 0) + 1
return {k: round(totals[k] / counts[k], 4) for k in sorted(totals)}
@staticmethod
def _bar(value: float, width: int = 20) -> str:
"""ASCII progress bar."""
filled = int(value * width)
return "[" + "#" * filled + "." * (width - filled) + "]"
# -- save / load results -----------------------------------------------
def save_results(self, results: Dict[str, Any], filepath: str) -> None:
"""Save evaluation results to JSON."""
# Convert non-serialisable types
os.makedirs(os.path.dirname(filepath) or ".", exist_ok=True)
with open(filepath, "w", encoding="utf-8") as f:
json.dump(results, f, indent=2, default=str)
@staticmethod
def load_results(filepath: str) -> Dict[str, Any]:
"""Load evaluation results from JSON."""
with open(filepath, "r", encoding="utf-8") as f:
return json.load(f)
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def main() -> None:
parser = argparse.ArgumentParser(
description="Codette Benchmark Runner - evaluate model reasoning quality"
)
parser.add_argument(
"--responses", "-r",
required=True,
help="Path to JSON file with pre-generated responses (prompt -> response)",
)
parser.add_argument(
"--prompts-dir", "-p",
default=None,
help="Directory containing prompt JSON files (default: evaluation/prompts/)",
)
parser.add_argument(
"--baseline", "-b",
default=None,
help="Path to baseline responses JSON for comparison",
)
parser.add_argument(
"--output", "-o",
default=None,
help="Save results to this JSON file",
)
parser.add_argument(
"--counterexamples", "-c",
action="store_true",
help="Also run counterexample tests",
)
parser.add_argument(
"--prompts-file",
default="reasoning_tests.json",
help="Prompt file name inside prompts dir (default: reasoning_tests.json)",
)
args = parser.parse_args()
runner = BenchmarkRunner(prompts_dir=args.prompts_dir)
runner.load_prompts(args.prompts_file)
print(f"Loading responses from: {args.responses}")
responses = runner.load_responses(args.responses)
print(f" Loaded {len(responses)} responses")
# Score
print("\nScoring responses...")
results = runner.score_responses(responses)
print(runner.format_report(results))
# Counterexamples
if args.counterexamples:
print("\nRunning counterexample tests...")
runner.load_counterexamples()
ce_results = runner.score_counterexamples(responses)
print(f" Refutation detection rate: {ce_results['refutation_rate']:.2%}")
results["counterexamples"] = ce_results
# Comparison
if args.baseline:
print(f"\nLoading baseline from: {args.baseline}")
baseline = runner.load_responses(args.baseline)
comparison = runner.compare_models(baseline, responses)
print(runner.format_comparison_report(comparison))
results["comparison"] = comparison
# Save
if args.output:
runner.save_results(results, args.output)
print(f"\nResults saved to: {args.output}")
if __name__ == "__main__":
main()
|