File size: 16,291 Bytes
96abbd8 | 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 | """
Build a debate-specific memory bank from historical debate runs.
This scans existing debate result directories such as
`./results/Agora-Opt/debate/<dataset>/<timestamp>_<modelA>_vs_<modelB>/`
directories, identifies problems where the two single generations disagreed yet
the debate converged to a correct consensus, summarizes the key reconciliation
insights (optionally via an LLM), and stores the cases inside a dedicated
`MemoryBank` directory (default: ./debate_memory_storage).
"""
from __future__ import annotations
import argparse
import json
import math
import os
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, Iterable, List, Optional, Tuple
from tqdm import tqdm
from .llm import get_response
from .memory_bank import MemoryBank
PKG_DIR = Path(__file__).resolve().parent
PROJECT_ROOT = PKG_DIR.parent.parent
DEFAULT_RUNS_ROOT = PROJECT_ROOT.parent.parent / "results" / "Agora-Opt" / "debate"
DEFAULT_DEBATE_MEMORY_DIR = PROJECT_ROOT / "debate_memory_storage"
@dataclass
class DebateCaseInput:
dataset: str
problem_id: int
description: str
final_code: str
final_result: Optional[float]
debate_rounds: List[Dict]
modelA: str
modelB: str
run_dir: Path
ground_truth: Optional[str]
initial_A_result: Optional[float]
initial_B_result: Optional[float]
evaluation: Dict
metadata: Dict
def load_jsonl(path: Path) -> List[Dict]:
if not path.exists():
return []
data: List[Dict] = []
with path.open("r", encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if not line:
continue
try:
data.append(json.loads(line))
except json.JSONDecodeError:
continue
return data
def float_or_none(value) -> Optional[float]:
if value is None:
return None
try:
return float(value)
except (ValueError, TypeError):
return None
def has_disagreement(entry: Dict, tolerance: float) -> bool:
a = float_or_none(entry.get("initial_A_result"))
b = float_or_none(entry.get("initial_B_result"))
if a is None or b is None:
return True
return abs(a - b) > tolerance
def summarize_rounds(rounds: List[Dict], max_chars: int = 2000) -> str:
if not rounds:
return ""
lines: List[str] = []
for rnd in rounds:
round_idx = rnd.get("round")
res_a = rnd.get("result_A")
res_b = rnd.get("result_B")
status_a = rnd.get("status_A")
status_b = rnd.get("status_B")
analysis_a = (rnd.get("analysis_A") or "").strip()
analysis_b = (rnd.get("analysis_B") or "").strip()
lines.append(
f"Round {round_idx}: A={res_a} ({status_a}), B={res_b} ({status_b})"
)
if analysis_a:
lines.append(f"Model A analysis:\n{analysis_a}")
if analysis_b:
lines.append(f"Model B analysis:\n{analysis_b}")
lines.append("")
text = "\n".join(lines).strip()
if len(text) <= max_chars:
return text
return text[: max_chars - 200] + "\n...\n(truncated)"
def build_summary_payload(
case: DebateCaseInput,
llm_model: Optional[str],
temperature: float,
llm_attempts: int = 1,
) -> Dict:
history_text = summarize_rounds(case.debate_rounds)
default_summary = {
"summary": (
f"Initial mismatch: modelA={case.initial_A_result}, "
f"modelB={case.initial_B_result}. "
f"Debate converged in {len(case.debate_rounds)} rounds."
),
"mismatch_reason": "",
"decisive_argument": "",
"guardrails": [],
"modeling_patterns": [],
}
if not llm_model:
return default_summary | {"history_excerpt": history_text}
prompt = f"""
You are helping an optimisation-debate memory builder.
Problem description:
{case.description}
Initial disagreement:
- Model A result: {case.initial_A_result}
- Model B result: {case.initial_B_result}
- Ground truth (if known): {case.ground_truth}
Debate transcript:
{history_text}
Final consensus objective: {case.final_result}
Please return a JSON object with the following keys:
- "summary": 2-3 sentences explaining how the debate resolved the mismatch.
- "mismatch_reason": concise reason for the disagreement.
- "decisive_argument": specific insight that convinced both sides.
- "guardrails": list of actionable bullet points the next debater should follow.
- "modeling_patterns": list of reusable modeling tricks/structures that appeared.
JSON ONLY. No prose outside the JSON.
""".strip()
attempts_remaining = max(1, llm_attempts)
last_error: Optional[Exception] = None
while attempts_remaining > 0:
try:
response = get_response(
prompt,
model=llm_model,
temperature=temperature,
maximum_retries=1,
)
payload = json.loads(response)
payload["history_excerpt"] = history_text
return payload
except Exception as exc: # noqa: BLE001
last_error = exc
attempts_remaining -= 1
fallback = default_summary.copy()
failure_reason = f"{last_error}" if last_error else "LLM call failed"
fallback["summary"] += f" LLM summary failed: {failure_reason}"
fallback["history_excerpt"] = history_text
return fallback
def existing_signatures(memory_dir: Path) -> set[str]:
cases_path = memory_dir / "cases.jsonl"
if not cases_path.exists():
return set()
signs: set[str] = set()
with cases_path.open("r", encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if not line:
continue
try:
data = json.loads(line)
except json.JSONDecodeError:
continue
meta = data.get("metadata") or {}
sig = meta.get("debate_signature")
if sig:
signs.add(sig)
return signs
class DebateMemoryBuilder:
def __init__(
self,
runs_root: Path,
memory_dir: Path,
mismatch_tolerance: float,
llm_model: Optional[str],
temperature: float,
llm_attempts: int,
max_workers: int,
datasets: Optional[Iterable[str]] = None,
dry_run: bool = False,
) -> None:
self.runs_root = runs_root
self.memory_dir = memory_dir
self.mismatch_tolerance = mismatch_tolerance
self.llm_model = llm_model
self.temperature = temperature
self.llm_attempts = max(1, llm_attempts)
self.max_workers = max_workers
self.datasets_filter = {d.lower() for d in datasets} if datasets else None
self.dry_run = dry_run
def build(self) -> None:
candidates = self._collect_candidates()
if not candidates:
print("No qualifying debate cases found.")
return
if not self.memory_dir.exists() and not self.dry_run:
self.memory_dir.mkdir(parents=True, exist_ok=True)
seen_sigs = existing_signatures(self.memory_dir)
bank = None if self.dry_run else MemoryBank(memory_dir=str(self.memory_dir))
added = 0
skipped_duplicates = 0
with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
futures = {
executor.submit(self._summarize_case, case): case
for case in candidates
}
for future in tqdm(as_completed(futures), total=len(futures), desc="Summarizing"):
case = futures[future]
signature = f"{case.dataset}:{case.problem_id}:{case.run_dir.name}"
if signature in seen_sigs:
skipped_duplicates += 1
continue
summary_payload = future.result()
description = (
f"{case.description.strip()}\n\n"
f"# Debate Memory Summary\n"
f"{summary_payload.get('summary', '').strip()}"
).strip()
metadata = {
"source": "debate_memory_builder",
"dataset": case.dataset,
"run_dir": str(case.run_dir),
"modelA": case.modelA,
"modelB": case.modelB,
"initial_A_result": case.initial_A_result,
"initial_B_result": case.initial_B_result,
"ground_truth": case.ground_truth,
"debate_signature": signature,
"summary": summary_payload,
}
if self.dry_run:
added += 1
continue
try:
bank.add_case(
problem_id=case.problem_id,
problem_desc=description,
solution_code=case.final_code,
objective_value=case.final_result or 0.0,
is_correct=True,
metadata=metadata,
)
added += 1
seen_sigs.add(signature)
except Exception as exc: # noqa: BLE001
print(f"Failed to add case {signature}: {exc}")
print("===== Debate Memory Builder Summary =====")
print(f"Runs root: {self.runs_root}")
print(f"Output dir: {self.memory_dir}")
print(f"Total candidates: {len(candidates)}")
print(f"Added cases: {added}")
print(f"Duplicates skipped: {skipped_duplicates}")
if self.dry_run:
print("Dry-run mode: no cases were written.")
def _collect_candidates(self) -> List[DebateCaseInput]:
candidates: List[DebateCaseInput] = []
if not self.runs_root.exists():
print(f"Runs root not found: {self.runs_root}")
return candidates
for dataset_dir in sorted(self.runs_root.iterdir()):
if not dataset_dir.is_dir():
continue
dataset_name = dataset_dir.name
if self.datasets_filter and dataset_name.lower() not in self.datasets_filter:
continue
for run_dir in sorted(dataset_dir.iterdir()):
if not run_dir.is_dir():
continue
dataset_candidates = self._parse_run(dataset_name, run_dir)
candidates.extend(dataset_candidates)
return candidates
def _parse_run(self, dataset: str, run_dir: Path) -> List[DebateCaseInput]:
results_path = run_dir / "debate_results.jsonl"
if not results_path.exists():
return []
modelA, modelB = self._infer_models(run_dir.name)
consensus_path = next(run_dir.glob("consensus_*_vs_*.jsonl"), None)
consensus_records = load_jsonl(consensus_path) if consensus_path else []
desc_map = {int(rec["id"]): rec for rec in consensus_records if "id" in rec}
eval_path = run_dir / "eval_consensus" / "evaluation_results.jsonl"
evaluation_map = {
int(rec["id"]): rec for rec in load_jsonl(eval_path) if "id" in rec
}
run_candidates: List[DebateCaseInput] = []
for entry in load_jsonl(results_path):
problem_id = entry.get("problem_id")
if problem_id is None:
continue
problem_id = int(problem_id)
if not has_disagreement(entry, self.mismatch_tolerance):
continue
if not entry.get("converged"):
continue
evaluation = evaluation_map.get(problem_id)
desc_entry = desc_map.get(problem_id)
if desc_entry:
description = desc_entry.get("description") or f"{dataset} problem {problem_id}"
else:
description = f"Dataset {dataset} problem {problem_id}"
final_code = entry.get("final_code") or (
desc_entry.get("generated_code", "") if desc_entry else ""
)
if not final_code:
continue
debate_rounds = entry.get("debate_rounds") or []
if not debate_rounds:
continue
run_candidates.append(
DebateCaseInput(
dataset=dataset,
problem_id=problem_id,
description=description,
final_code=final_code,
final_result=float_or_none(entry.get("final_result")),
debate_rounds=debate_rounds,
modelA=modelA,
modelB=modelB,
run_dir=run_dir,
ground_truth=entry.get("ground_truth"),
initial_A_result=float_or_none(entry.get("initial_A_result")),
initial_B_result=float_or_none(entry.get("initial_B_result")),
evaluation=evaluation or {},
metadata={
"run_dir": str(run_dir),
"dataset": dataset,
},
)
)
return run_candidates
@staticmethod
def _infer_models(run_name: str) -> Tuple[str, str]:
"""
Run folder format: <timestamp>_<modelA>_vs_<modelB>
"""
parts = run_name.split("_vs_")
if len(parts) != 2:
return "modelA", "modelB"
left = parts[0].split("_") # timestamp + modelA pieces
if len(left) < 2:
return left[-1], parts[1]
modelA = "_".join(left[1:])
modelB = parts[1]
return modelA, modelB
def _summarize_case(self, case: DebateCaseInput) -> Dict:
return build_summary_payload(
case,
llm_model=self.llm_model,
temperature=self.temperature,
llm_attempts=self.llm_attempts,
)
def parse_args():
parser = argparse.ArgumentParser(description="Build debate memory bank from historical runs.")
parser.add_argument(
"--runs_root",
type=str,
default=str(DEFAULT_RUNS_ROOT),
help="Directory containing debate run artifacts.",
)
parser.add_argument(
"--output_dir",
type=str,
default=str(DEFAULT_DEBATE_MEMORY_DIR),
help="Directory to store the debate memory bank.",
)
parser.add_argument(
"--datasets",
type=str,
nargs="*",
default=None,
help="Optional dataset filters (case-insensitive).",
)
parser.add_argument(
"--mismatch_tolerance",
type=float,
default=1e-3,
help="Minimum absolute difference between initial results to consider a disagreement.",
)
parser.add_argument(
"--llm_model",
type=str,
default=None,
help="Optional model name for LLM-based summaries. If omitted, heuristic summaries are used.",
)
parser.add_argument(
"--temperature",
type=float,
default=0.3,
help="Temperature for LLM summaries.",
)
parser.add_argument(
"--max_workers",
type=int,
default=4,
help="Parallel workers for summary generation.",
)
parser.add_argument(
"--llm_attempts",
type=int,
default=2,
help="Number of LLM attempts per case before falling back to heuristics.",
)
parser.add_argument(
"--dry_run",
action="store_true",
help="Run the pipeline without writing to the memory bank.",
)
return parser.parse_args()
def main():
args = parse_args()
builder = DebateMemoryBuilder(
runs_root=Path(args.runs_root),
memory_dir=Path(args.output_dir),
mismatch_tolerance=args.mismatch_tolerance,
llm_model=args.llm_model,
temperature=args.temperature,
llm_attempts=args.llm_attempts,
max_workers=args.max_workers,
datasets=args.datasets,
dry_run=args.dry_run,
)
builder.build()
if __name__ == "__main__":
main()
|