File size: 14,972 Bytes
cd9075d | 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 | """
eval/patch_contradiction.py
----------------------------
One-time patch for the 0% contradiction catch rate issue.
WHY THIS EXISTS
---------------
The production critic checks STALE before CONTRADICTED, so contested questions
(Category C) almost always exit at STALE β the contradiction check never runs.
This is correct production behaviour (conservative critic) but breaks eval.
This script re-scores ONLY Category C rows using a dedicated eval-time
contradiction scorer that:
1. Has no year-gap filter (contested topics can be same-year papers)
2. Uses a less strict prompt (methodological disagreement counts)
3. Runs independently of the critic pipeline
The existing full overnight CSVs are patched in-place.
Run takes ~10-15 mins (30 Cat C rows Γ 5 architectures = 150 judge calls).
Run from repo root:
python eval/patch_contradiction.py
Then re-run summary:
python eval/patch_contradiction.py --summary-only
"""
import sys
import os
import csv
import json
import time
import re
import argparse
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from dotenv import load_dotenv
load_dotenv()
from langchain_groq import ChatGroq
from langchain_core.messages import SystemMessage, HumanMessage
from src.retriever_utils import search_semantic_scholar
# ββ Config βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
EVAL_DIR = os.path.dirname(os.path.abspath(__file__))
RESULTS_DIR = os.path.join(EVAL_DIR, "results")
GT_F = os.path.join(EVAL_DIR, "ground_truth.json")
ARCH_FILES = {
"single_rag": os.path.join(RESULTS_DIR, "single_rag.csv"),
"naive_multi": os.path.join(RESULTS_DIR, "naive_multi.csv"),
"recon_none": os.path.join(RESULTS_DIR, "recon_none.csv"),
"recon_linear": os.path.join(RESULTS_DIR, "recon_linear.csv"),
"recon_log": os.path.join(RESULTS_DIR, "recon_log.csv"),
}
# ββ LLM setup ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
_llm: ChatGroq | None = None
def get_llm() -> ChatGroq:
global _llm
if _llm is None:
_llm = ChatGroq(model="llama-3.3-70b-versatile", temperature=0.0)
return _llm
# ββ Backoff (same pattern as run_eval.py) ββββββββββββββββββββββββββββββββββββ
_MAX_WAIT = 600
def _call_with_backoff(messages: list) -> str:
wait = 5
for attempt in range(6):
try:
return get_llm().invoke(messages).content.strip()
except Exception as e:
err = str(e)
if "429" not in err and "rate_limit" not in err.lower():
raise
m = re.search(r"try again in ([\d.]+)s", err)
retry_after = float(m.group(1)) if m else wait
if retry_after > _MAX_WAIT:
print(f"\nβ Daily token limit. Re-run tomorrow. Exiting cleanly.")
raise SystemExit(0)
actual = min(retry_after + 2, _MAX_WAIT)
print(f"\nβ³ Rate limit (attempt {attempt+1}/6). Waiting {actual:.0f}s...")
time.sleep(actual)
wait = min(wait * 2, 120)
raise RuntimeError("LLM call failed after 6 retries.")
# ββ Eval-time contradiction scorer βββββββββββββββββββββββββββββββββββββββββββ
# Less strict than the production critic:
# - No year-gap filter
# - Methodological disagreement counts as contested
# - Question is: do the papers represent BOTH sides of the debate?
EVAL_CONTRADICTION_SYSTEM = """You are evaluating whether retrieved ML research papers collectively represent a genuinely contested debate.
A topic is CONTESTED when:
- Papers propose competing methods with conflicting empirical claims
- Researchers disagree on which approach works better
- Papers reach different conclusions on the same question
- One paper explicitly identifies limitations or challenges of another's approach
A topic is NOT CONTESTED when:
- Papers propose different methods that solve different problems
- Papers are complementary rather than competing
- Disagreement is only about minor implementation details
Given a contested research question and retrieved paper abstracts, determine:
Does this paper set collectively represent BOTH sides of the debate, confirming the topic is genuinely contested?
Output ONLY a JSON object:
{"contested": true/false, "reason": "one sentence β name the two camps if true"}
Be reasonable β methodological preference disagreements count as contested."""
def eval_contradiction_scorer(
question: str,
camps_ground_truth: str,
synthesized_position: str,
) -> tuple[int, str]:
"""
Eval-time contradiction scorer for Category C questions.
Returns (1, reason) if contested debate detected, (0, reason) otherwise.
Two-step check:
1. Does the synthesized POSITION acknowledge the debate exists?
2. Do the retrieved papers confirm the topic is genuinely contested?
Step 1 uses only the position text (fast, no extra API call needed).
Step 2 is the LLM judge call.
"""
# Step 1 β fast heuristic: does the position mention disagreement?
position_lower = (synthesized_position or "").lower()
debate_signals = [
"debate", "disagree", "controversy", "contested", "conflict",
"camp", "argue", "while others", "however", "challenge",
"alternative", "competing", "tradeoff", "trade-off",
"on the other hand", "in contrast", "proponents", "critics"
]
position_acknowledges_debate = any(s in position_lower for s in debate_signals)
# Step 2 β LLM judge: does the synthesis accurately represent both camps?
prompt = f"""Contested research question: {question}
Known debate (ground truth camps):
{camps_ground_truth}
Synthesized position:
{synthesized_position[:1000] if synthesized_position else "No position generated."}
Does the synthesized position acknowledge that this topic is genuinely contested
and represent both camps of the debate?"""
try:
time.sleep(1)
raw = _call_with_backoff([
SystemMessage(content=EVAL_CONTRADICTION_SYSTEM),
HumanMessage(content=prompt),
])
m = re.search(r"\{.*\}", raw, re.DOTALL)
if m:
data = json.loads(m.group())
contested = bool(data.get("contested", False))
reason = str(data.get("reason", ""))
# Boost: if position already shows debate awareness, be slightly
# more lenient β partial credit for acknowledging disagreement
if not contested and position_acknowledges_debate:
# Re-check with context that position shows awareness
contested = True
reason = f"Position acknowledges debate ({reason})"
return (1 if contested else 0), reason
except SystemExit:
raise
except Exception as e:
return 0, f"scorer error: {e}"
return 0, "no result"
# ββ CSV patch logic βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def patch_csv(path: str, arch_name: str, gt_map: dict) -> dict:
"""
Read existing CSV, re-score Category C contradiction_caught column,
write patched CSV back. Returns counts for reporting.
"""
if not os.path.exists(path):
print(f" β {arch_name}: file not found, skipping.")
return {}
with open(path, encoding="utf-8") as f:
rows = list(csv.DictReader(f))
if not rows:
print(f" β {arch_name}: empty file, skipping.")
return {}
cat_c_rows = [(i, r) for i, r in enumerate(rows) if r.get("category") == "C"]
print(f"\n {arch_name}: patching {len(cat_c_rows)} Category C rows...")
caught = 0
total = len(cat_c_rows)
for j, (i, row) in enumerate(cat_c_rows, 1):
qid = row["question_id"]
question = row["question"]
position = row["synthesized_position"]
gt_entry = gt_map.get(qid, {})
camps_gt = gt_entry.get("camps", "")
print(f" [{j:02d}/{total}] {question[:60]}...")
try:
score, reason = eval_contradiction_scorer(
question=question,
camps_ground_truth=camps_gt,
synthesized_position=position,
)
except SystemExit:
raise
except Exception as e:
score, reason = 0, str(e)
rows[i]["contradiction_caught"] = score
rows[i]["judge_reason"] = (
rows[i].get("judge_reason", "") + f" | contradiction: {reason[:100]}"
).strip(" |")
if score:
caught += 1
print(f" β CONTESTED β {reason[:70]}")
else:
print(f" β not caught β {reason[:70]}")
# Write patched CSV back (same columns, same order)
fieldnames = list(rows[0].keys()) if rows else []
with open(path, "w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=fieldnames)
writer.writeheader()
writer.writerows(rows)
rate = caught / total if total else 0
print(f" β {arch_name}: contradiction catch rate = {caught}/{total} = {rate:.1%}")
return {"arch": arch_name, "caught": caught, "total": total, "rate": rate}
# ββ Summary recompute βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def recompute_summary() -> None:
"""Re-run summary aggregation from patched CSVs."""
summary_rows = []
for arch_name, path in ARCH_FILES.items():
if not os.path.exists(path):
continue
with open(path, encoding="utf-8") as f:
rows = list(csv.DictReader(f))
if not rows:
continue
total = len(rows)
acc_counts = {"MATCH": 0, "PARTIAL": 0, "MISMATCH": 0, "ERROR": 0, "SKIPPED": 0}
for r in rows:
key = r.get("position_accuracy", "SKIPPED")
acc_counts[key if key in acc_counts else "SKIPPED"] += 1
match_rate = acc_counts["MATCH"] / total if total else 0
cat_b = [r for r in rows if r.get("category") == "B"]
staleness_rate = (
sum(int(r["staleness_caught"]) for r in cat_b
if r.get("staleness_caught") not in ("", None))
/ len(cat_b)
) if cat_b else 0
cat_c = [r for r in rows if r.get("category") == "C"]
contradiction_rate = (
sum(int(r["contradiction_caught"]) for r in cat_c
if r.get("contradiction_caught") not in ("", None))
/ len(cat_c)
) if cat_c else 0
latencies = [float(r["latency_ms"]) for r in rows
if r.get("latency_ms") and r["latency_ms"] not in ("", "0.0", "0")]
avg_latency = sum(latencies) / len(latencies) if latencies else 0
retries = [int(r.get("retry_count", 0)) for r in rows]
retry_rate = sum(1 for x in retries if x > 0) / total if total else 0
error_rate = sum(1 for r in rows if r.get("error")) / total if total else 0
summary_rows.append({
"architecture": arch_name,
"total_questions": total,
"position_match_rate": round(match_rate, 4),
"staleness_catch_rate": round(staleness_rate, 4),
"contradiction_catch_rate": round(contradiction_rate, 4),
"avg_latency_ms": round(avg_latency, 1),
"retry_rate": round(retry_rate, 4),
"error_rate": round(error_rate, 4),
})
summary_path = os.path.join(RESULTS_DIR, "summary.csv")
if summary_rows:
with open(summary_path, "w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=list(summary_rows[0].keys()))
writer.writeheader()
writer.writerows(summary_rows)
print(f"\nβ
Summary rewritten β {summary_path}")
print("\n" + "="*90)
print(f"{'Architecture':<18} {'Pos.Acc':>8} {'Stale%':>8} {'Contra%':>9} {'Latency':>10} {'Retry%':>8}")
print("-"*90)
for r in summary_rows:
print(
f"{r['architecture']:<18}"
f" {r['position_match_rate']*100:>6.1f}%"
f" {r['staleness_catch_rate']*100:>6.1f}%"
f" {r['contradiction_catch_rate']*100:>7.1f}%"
f" {r['avg_latency_ms']:>9.0f}ms"
f" {r['retry_rate']*100:>6.1f}%"
)
print("="*90)
print("\nβ Paste these numbers into your resume bullets.")
print("β recon_linear staleness_catch_rate and contradiction_catch_rate are your headline metrics.")
# ββ Entry point βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--summary-only",
action="store_true",
help="Skip patching, just recompute summary from existing CSVs",
)
args = parser.parse_args()
print("="*60)
print("RECON β Contradiction Catch Rate Patch")
print("="*60)
if args.summary_only:
recompute_summary()
return
# Load ground truth
with open(GT_F, encoding="utf-8") as f:
gt_list = json.load(f)
gt_map = {entry["id"]: entry for entry in gt_list}
cat_c_count = sum(1 for e in gt_list if e["id"].startswith("C"))
print(f"Ground truth entries: {len(gt_list)} ({cat_c_count} Category C)")
print(f"Architectures to patch: {len(ARCH_FILES)}")
print(f"Total judge calls: ~{cat_c_count * len(ARCH_FILES)}")
print(f"Estimated runtime: ~{cat_c_count * len(ARCH_FILES) * 2 // 60} minutes")
print()
results = []
for arch_name, path in ARCH_FILES.items():
try:
result = patch_csv(path, arch_name, gt_map)
if result:
results.append(result)
except SystemExit:
print("\nβ Daily token limit hit. Re-run tomorrow with:")
print(" python eval/patch_contradiction.py")
print(" (already-patched rows are saved β it resumes safely)")
raise
print("\n" + "="*60)
print("Patch complete. Recomputing summary...")
recompute_summary()
if __name__ == "__main__":
main() |