File size: 14,618 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 | """
Failure Analyzer - examines evaluation logs to find patterns in
low-scoring responses, cluster failures by topic, and recommend
dataset improvements.
"""
from __future__ import annotations
import json
import re
import sys
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any, Dict, List, Optional, Set, Tuple
_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))
# ---------------------------------------------------------------------------
# Keyword extraction (lightweight, no external deps)
# ---------------------------------------------------------------------------
_STOP_WORDS: Set[str] = {
"the", "a", "an", "is", "are", "was", "were", "be", "been", "being",
"have", "has", "had", "do", "does", "did", "will", "would", "shall",
"should", "may", "might", "must", "can", "could", "to", "of", "in",
"for", "on", "with", "at", "by", "from", "as", "into", "through",
"during", "before", "after", "above", "below", "between", "out",
"off", "over", "under", "again", "further", "then", "once", "here",
"there", "when", "where", "why", "how", "all", "both", "each",
"few", "more", "most", "other", "some", "such", "no", "nor", "not",
"only", "own", "same", "so", "than", "too", "very", "just", "don",
"now", "and", "but", "or", "if", "while", "that", "this", "what",
"which", "who", "whom", "it", "its", "they", "them", "their",
"he", "she", "him", "her", "his", "we", "us", "our", "you", "your",
"i", "me", "my", "about", "up",
}
def _extract_keywords(text: str, top_n: int = 8) -> List[str]:
"""Extract the most frequent meaningful words from text."""
words = re.findall(r"[a-z]{3,}", text.lower())
filtered = [w for w in words if w not in _STOP_WORDS]
counts = Counter(filtered)
return [w for w, _ in counts.most_common(top_n)]
def _jaccard(set_a: Set[str], set_b: Set[str]) -> float:
"""Jaccard similarity between two sets."""
if not set_a and not set_b:
return 1.0
union = set_a | set_b
if not union:
return 0.0
return len(set_a & set_b) / len(union)
# ---------------------------------------------------------------------------
# FailureAnalyzer
# ---------------------------------------------------------------------------
class FailureAnalyzer:
"""Analyze evaluation results to identify failure patterns."""
# Score thresholds
FAILURE_THRESHOLD = 0.4 # scores below this = failure
WEAK_THRESHOLD = 0.55 # scores below this = weak
def __init__(
self,
failure_threshold: float = 0.4,
weak_threshold: float = 0.55,
):
self.failure_threshold = failure_threshold
self.weak_threshold = weak_threshold
# -- loading -----------------------------------------------------------
@staticmethod
def load_results(filepath: str) -> Dict[str, Any]:
"""Load benchmark results JSON produced by BenchmarkRunner."""
with open(filepath, "r", encoding="utf-8") as f:
return json.load(f)
# -- analysis ----------------------------------------------------------
def find_failures(
self,
results: Dict[str, Any],
dimension: str = "overall",
) -> List[Dict[str, Any]]:
"""Return entries whose *dimension* score is below failure threshold."""
failures = []
for entry in results.get("all_scores", []):
score = entry.get("scores", {}).get(dimension)
if score is not None and score < self.failure_threshold:
failures.append({
"prompt": entry["prompt"],
"score": score,
"all_scores": entry["scores"],
})
failures.sort(key=lambda x: x["score"])
return failures
def find_weak_areas(
self,
results: Dict[str, Any],
) -> Dict[str, float]:
"""Identify which scoring dimensions are weakest across all prompts.
Returns dict of dimension -> average score, sorted ascending.
"""
dimension_totals: Dict[str, float] = defaultdict(float)
dimension_counts: Dict[str, int] = defaultdict(int)
for entry in results.get("all_scores", []):
for k, v in entry.get("scores", {}).items():
if isinstance(v, float) and k not in ("word_count", "sentence_count"):
dimension_totals[k] += v
dimension_counts[k] += 1
averages = {}
for k in dimension_totals:
if dimension_counts[k] > 0:
averages[k] = round(dimension_totals[k] / dimension_counts[k], 4)
return dict(sorted(averages.items(), key=lambda x: x[1]))
def failure_rate_by_category(
self,
results: Dict[str, Any],
dimension: str = "overall",
) -> Dict[str, Dict[str, Any]]:
"""Calculate failure rates per category."""
rates: Dict[str, Dict[str, Any]] = {}
for cat, data in results.get("categories", {}).items():
details = data.get("details", [])
total = len(details)
if total == 0:
continue
failures = sum(
1 for d in details
if d.get("scores", {}).get(dimension, 1.0) < self.failure_threshold
)
weak = sum(
1 for d in details
if self.failure_threshold <= d.get("scores", {}).get(dimension, 1.0) < self.weak_threshold
)
rates[cat] = {
"total": total,
"failures": failures,
"weak": weak,
"failure_rate": round(failures / total, 4),
"weak_rate": round(weak / total, 4),
"avg_score": data.get("average_scores", {}).get(dimension, 0),
}
return dict(sorted(rates.items(), key=lambda x: -x[1]["failure_rate"]))
def cluster_failures_by_topic(
self,
failures: List[Dict[str, Any]],
similarity_threshold: float = 0.25,
) -> List[Dict[str, Any]]:
"""Cluster failure prompts by keyword overlap.
Uses a simple greedy clustering: each prompt is assigned to the first
cluster whose centroid keywords have Jaccard similarity above threshold.
"""
clusters: List[Dict[str, Any]] = []
for failure in failures:
prompt = failure["prompt"]
keywords = set(_extract_keywords(prompt))
matched = False
for cluster in clusters:
if _jaccard(keywords, cluster["keywords"]) >= similarity_threshold:
cluster["prompts"].append(failure)
cluster["keywords"] |= keywords
matched = True
break
if not matched:
clusters.append({
"keywords": keywords,
"prompts": [failure],
})
# Format output
result = []
for i, c in enumerate(clusters):
avg_score = sum(p["score"] for p in c["prompts"]) / len(c["prompts"])
result.append({
"cluster_id": i,
"topic_keywords": sorted(c["keywords"])[:10],
"num_failures": len(c["prompts"]),
"avg_score": round(avg_score, 4),
"sample_prompts": [p["prompt"] for p in c["prompts"][:5]],
})
result.sort(key=lambda x: -x["num_failures"])
return result
def identify_weakest_dimensions(
self,
results: Dict[str, Any],
top_n: int = 3,
) -> List[Tuple[str, float]]:
"""Return the top_n weakest scoring dimensions."""
averages = self.find_weak_areas(results)
items = [(k, v) for k, v in averages.items() if k != "overall"]
return items[:top_n]
# -- recommendations ---------------------------------------------------
def generate_recommendations(
self,
results: Dict[str, Any],
) -> List[str]:
"""Generate actionable recommendations for dataset improvement."""
recommendations: List[str] = []
# Weakest dimensions
weakest = self.identify_weakest_dimensions(results, top_n=3)
for dim, score in weakest:
if score < self.failure_threshold:
recommendations.append(
f"CRITICAL: Dimension '{dim}' averages {score:.3f} (below failure threshold). "
f"Add training examples that emphasise {dim} explicitly."
)
elif score < self.weak_threshold:
recommendations.append(
f"IMPROVE: Dimension '{dim}' averages {score:.3f} (weak). "
f"Augment dataset with responses demonstrating strong {dim}."
)
# Category failure rates
cat_rates = self.failure_rate_by_category(results)
for cat, info in cat_rates.items():
if info["failure_rate"] > 0.3:
recommendations.append(
f"CATEGORY '{cat}': {info['failure_rate']:.0%} failure rate. "
f"Add more diverse training examples for {cat} topics."
)
# Failure clustering
failures = self.find_failures(results)
if failures:
clusters = self.cluster_failures_by_topic(failures)
for cluster in clusters[:3]:
kw = ", ".join(cluster["topic_keywords"][:5])
recommendations.append(
f"TOPIC CLUSTER: {cluster['num_failures']} failures around "
f"[{kw}]. Create targeted training data for these concepts."
)
# General
overall = results.get("overall", {})
overall_score = overall.get("overall", 0)
if overall_score < 0.5:
recommendations.append(
"GENERAL: Overall score is very low. Consider increasing dataset size "
"and diversity before next training run."
)
elif overall_score < 0.65:
recommendations.append(
"GENERAL: Overall score is moderate. Focus on the weakest categories "
"and dimensions for the next dataset iteration."
)
if not recommendations:
recommendations.append(
"No critical issues detected. Continue monitoring with additional benchmarks."
)
return recommendations
# -- report ------------------------------------------------------------
def format_report(self, results: Dict[str, Any]) -> str:
"""Generate a full failure analysis report."""
lines: List[str] = []
lines.append("=" * 70)
lines.append(" FAILURE ANALYSIS REPORT")
lines.append("=" * 70)
# Weakest dimensions
lines.append("")
lines.append("-" * 70)
lines.append(" WEAKEST SCORING DIMENSIONS")
lines.append("-" * 70)
weak_areas = self.find_weak_areas(results)
for dim, score in list(weak_areas.items())[:6]:
status = "FAIL" if score < self.failure_threshold else (
"WEAK" if score < self.weak_threshold else "OK "
)
lines.append(f" [{status}] {dim:<22s} {score:.4f}")
# Category failure rates
lines.append("")
lines.append("-" * 70)
lines.append(" FAILURE RATES BY CATEGORY")
lines.append("-" * 70)
cat_rates = self.failure_rate_by_category(results)
for cat, info in cat_rates.items():
lines.append(
f" {cat:<18s} fail: {info['failure_rate']:>5.1%} "
f"weak: {info['weak_rate']:>5.1%} "
f"avg: {info['avg_score']:.4f}"
)
# Failure clusters
failures = self.find_failures(results)
if failures:
lines.append("")
lines.append("-" * 70)
lines.append(f" FAILURE CLUSTERS ({len(failures)} total failures)")
lines.append("-" * 70)
clusters = self.cluster_failures_by_topic(failures)
for c in clusters[:5]:
kw = ", ".join(c["topic_keywords"][:6])
lines.append(f" Cluster {c['cluster_id']}: "
f"{c['num_failures']} failures, "
f"avg score {c['avg_score']:.4f}")
lines.append(f" Topics: {kw}")
for p in c["sample_prompts"][:2]:
lines.append(f" - {p[:70]}...")
# Recommendations
lines.append("")
lines.append("-" * 70)
lines.append(" RECOMMENDATIONS")
lines.append("-" * 70)
recs = self.generate_recommendations(results)
for i, rec in enumerate(recs, 1):
lines.append(f" {i}. {rec}")
lines.append("")
lines.append("=" * 70)
return "\n".join(lines)
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def main() -> None:
import argparse
parser = argparse.ArgumentParser(
description="Codette Failure Analyzer - identify patterns in evaluation failures"
)
parser.add_argument(
"--results", "-r",
required=True,
help="Path to benchmark results JSON",
)
parser.add_argument(
"--failure-threshold", "-f",
type=float,
default=0.4,
help="Score threshold for failure (default: 0.4)",
)
parser.add_argument(
"--weak-threshold", "-w",
type=float,
default=0.55,
help="Score threshold for weak (default: 0.55)",
)
args = parser.parse_args()
analyzer = FailureAnalyzer(
failure_threshold=args.failure_threshold,
weak_threshold=args.weak_threshold,
)
results = analyzer.load_results(args.results)
print(analyzer.format_report(results))
if __name__ == "__main__":
main()
|