File size: 19,709 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 | #!/usr/bin/env python3
"""
Run debates between two models using memory-augmented single generations.
This script automatically locates the latest initial-solution files for the
specified models, runs the parallel debate workflow from `simple_rag/debate.py`,
and then evaluates the consensus solutions with `execute.py`.
Example:
python run_memory_debate.py \
--datasets ComplexLP EasyLP \
--max_rounds 3 \
--debate_workers 16 \
--execute_workers 128
"""
from __future__ import annotations
import argparse
import datetime as dt
import glob
import json
import os
import re
import subprocess
import sys
from pathlib import Path
from typing import Dict, List, Optional, Tuple
from .config import normalize_dataset_name
SCRIPT_DIR = Path(__file__).resolve().parent
SRC_DIR = SCRIPT_DIR.parent
PROJECT_ROOT = SRC_DIR.parent
MONOREPO_ROOT = PROJECT_ROOT.parent
STANDARD_RESULTS_ROOT = PROJECT_ROOT.parent.parent / "results" / "Agora-Opt"
DEFAULT_RESULTS_DIR = STANDARD_RESULTS_ROOT / "generation"
DEFAULT_OUTPUT_ROOT = STANDARD_RESULTS_ROOT / "debate"
DEFAULT_DEBATE_SCRIPT = MONOREPO_ROOT / "simple_rag" / "debate.py"
DEFAULT_EXECUTE_SCRIPT = PROJECT_ROOT / "scripts" / "execute.py"
DEFAULT_DEBATE_MEMORY_DIR = PROJECT_ROOT / "debate_memory_storage"
DEBATE_MEMORY_HEADER = "# Debate Memory Insights"
from .memory_bank import MemoryBank
def format_debate_memory_context(cases: List[Dict]) -> str:
if not cases:
return ""
lines = [DEBATE_MEMORY_HEADER, ""]
for idx, item in enumerate(cases, 1):
case = item["case"]
score = item.get("score", 0.0)
metadata = case.get("metadata", {})
dataset = metadata.get("dataset", "unknown")
summary = metadata.get("summary", {}).get("summary")
lines.append(f"## Case {idx} (similarity {score:.3f}, dataset {dataset})")
description = case.get("description", "").strip()
if description:
snippet = description if len(description) <= 800 else description[:800] + "\n..."
lines.append(snippet)
if summary:
lines.append("Summary: " + summary)
lines.append("---")
return "\n".join(lines).strip()
def build_debate_memory_contexts(
files: List[str],
debate_memory: MemoryBank,
dataset: str,
top_k: int,
) -> Dict[int, str]:
contexts: Dict[int, str] = {}
if debate_memory is None or top_k <= 0:
return contexts
for file_path in files:
with open(file_path, "r", encoding="utf-8") as fh:
for line in fh:
if not line.strip():
continue
data = json.loads(line)
problem_id = data.get("id")
if problem_id is None or problem_id in contexts:
continue
description = data.get("description", "")
if not description.strip():
contexts[problem_id] = ""
continue
cases = debate_memory.retrieve_similar_cases(
description,
top_k=top_k,
preferred_dataset=dataset,
)
contexts[problem_id] = format_debate_memory_context(cases)
return contexts
def maybe_enrich_generation_file(
source_path: str,
destination_path: str,
contexts: Dict[int, str],
) -> str:
if not contexts:
return source_path
changed = False
enriched_lines: List[str] = []
with open(source_path, "r", encoding="utf-8") as fh:
for line in fh:
if not line.strip():
continue
data = json.loads(line)
pid = data.get("id")
context = contexts.get(pid)
if context:
data["description"] = f"{data.get('description', '').strip()}\n\n{context}"
changed = True
enriched_lines.append(json.dumps(data, ensure_ascii=False))
if not changed:
return source_path
with open(destination_path, "w", encoding="utf-8") as fh:
for entry in enriched_lines:
fh.write(entry + "\n")
return destination_path
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Parallel debate runner for memory-enhanced single generations"
)
parser.add_argument(
"--modelA",
type=str,
default="gpt-4o",
help="First model in the debate (default: gpt-4o)",
)
parser.add_argument(
"--modelB",
type=str,
default="deepseek-chat",
help="Second model in the debate (default: deepseek-chat)",
)
parser.add_argument(
"--results_dir",
type=str,
default=str(DEFAULT_RESULTS_DIR),
help="Directory that stores initial-solution JSONL files",
)
parser.add_argument(
"--datasets",
nargs="*",
default=None,
help="Datasets to debate. If omitted, auto-detect common datasets.",
)
parser.add_argument(
"--output_root",
type=str,
default=str(DEFAULT_OUTPUT_ROOT),
help="Root directory to store debate/eval outputs",
)
parser.add_argument(
"--debate_script",
type=str,
default=str(DEFAULT_DEBATE_SCRIPT),
help="Path to simple_rag/debate.py (override if needed)",
)
parser.add_argument(
"--execute_script",
type=str,
default=str(DEFAULT_EXECUTE_SCRIPT),
help="Path to debate_with_memory/execute.py (override if needed)",
)
parser.add_argument(
"--max_rounds",
type=int,
default=3,
help="Maximum number of debate rounds (default: 3)",
)
parser.add_argument(
"--temperature",
type=float,
default=0.01,
help="Temperature for debate LLM calls (default: 0.01)",
)
parser.add_argument(
"--debate_workers",
type=int,
default=16,
help="Parallel workers for debate (ThreadPool inside debate.py)",
)
parser.add_argument(
"--execute_workers",
type=int,
default=128,
help="Parallel workers for execute.py evaluation",
)
parser.add_argument(
"--max_problems",
type=int,
default=None,
help="Optional cap on number of problems per dataset",
)
parser.add_argument(
"--tolerance",
type=float,
default=0.05,
help="Relative tolerance for evaluation accuracy comparison",
)
parser.add_argument(
"--timeout",
type=int,
default=90,
help="Timeout (seconds) for executing consensus code",
)
parser.add_argument(
"--relative_tolerance",
action="store_true",
help="Pass --use_relative_tolerance to execute.py",
)
parser.add_argument(
"--save_execution_stdout",
action="store_true",
help="Store stdout/stderr for consensus executions",
)
parser.add_argument(
"--execute_memory_dir",
type=str,
default=None,
help="Optional memory_storage directory forwarded to execute.py during consensus evaluation.",
)
parser.add_argument(
"--execute_debug_memory_path",
type=str,
default=None,
help="Optional debug_memory.jsonl path forwarded to execute.py during consensus evaluation.",
)
parser.add_argument(
"--execute_disable_debug_memory",
action="store_true",
help="Pass --disable_debug_memory to execute.py during consensus evaluation.",
)
parser.add_argument(
"--dry_run",
action="store_true",
help="Only print the planned actions without running debate/eval",
)
parser.add_argument(
"--debate_memory_dir",
type=str,
default=str(DEFAULT_DEBATE_MEMORY_DIR),
help="Directory containing debate memory cases for prompt augmentation",
)
parser.add_argument(
"--debate_memory_top_k",
type=int,
default=2,
help="How many debate memory cases to retrieve per problem",
)
parser.add_argument(
"--disable_debate_memory",
action="store_true",
help="Skip retrieval even if debate memory directory exists",
)
parser.add_argument(
"--embedding_model",
type=str,
default=None,
help="Embedding model name or local path (default: BAAI/bge-small-en-v1.5). "
"Use local path to avoid Hugging Face downloads, or set HF_HUB_OFFLINE=1 environment variable.",
)
return parser.parse_args()
def normalize_dataset_list(raw_list: Optional[List[str]]) -> Optional[List[str]]:
"""Split comma-separated values and strip whitespace."""
if not raw_list:
return None
datasets: List[str] = []
for item in raw_list:
parts = [part.strip() for part in item.split(",") if part.strip()]
datasets.extend(normalize_dataset_name(part) for part in parts)
return datasets or None
def collect_runs(results_dir: str, model: str) -> Dict[str, List[Tuple[str, str]]]:
"""
Return mapping dataset -> list of (timestamp, path) sorted ascending.
Skips evaluation artifacts (suffixes containing '_eval').
"""
pattern = os.path.join(results_dir, f"{model}_*.jsonl")
regex = re.compile(rf"{re.escape(model)}_(.+)_(\d{{8}}_\d{{6}})\.jsonl$")
runs: Dict[str, List[Tuple[str, str]]] = {}
for path in glob.glob(pattern):
base = os.path.basename(path)
match = regex.match(base)
if not match:
continue
dataset = normalize_dataset_name(match.group(1))
if "_eval" in dataset:
continue
timestamp = match.group(2)
runs.setdefault(dataset, []).append((timestamp, path))
for dataset in runs:
runs[dataset].sort() # chronological
return runs
def pick_latest(runs: Dict[str, List[Tuple[str, str]]], dataset: str) -> Optional[str]:
"""Return latest file path for dataset if available."""
entries = runs.get(dataset)
if not entries:
return None
return entries[-1][1]
def stream_command(cmd: List[str], cwd: str, log_path: str) -> None:
"""Run a subprocess, streaming output to stdout and a log file."""
print(f"\n▶ Running: {' '.join(cmd)}", flush=True)
print(f" cwd: {cwd}", flush=True)
os.makedirs(os.path.dirname(log_path), exist_ok=True)
with open(log_path, "w", encoding="utf-8") as log_file:
process = subprocess.Popen(
cmd,
cwd=cwd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
encoding="utf-8",
errors="replace",
bufsize=1,
)
assert process.stdout is not None # for type checkers
for line in process.stdout:
print(line, end="", flush=True)
log_file.write(line)
log_file.flush()
return_code = process.wait()
if return_code != 0:
raise subprocess.CalledProcessError(return_code, cmd)
def load_eval_report(report_path: str) -> Optional[Dict]:
if not os.path.exists(report_path):
return None
with open(report_path, "r", encoding="utf-8") as fh:
return json.load(fh)
def ensure_script(path: str, description: str) -> None:
if not os.path.isfile(path):
raise FileNotFoundError(f"{description} not found: {path}")
def main() -> None:
args = parse_args()
args.datasets = normalize_dataset_list(args.datasets)
args.output_root = os.path.abspath(args.output_root)
args.results_dir = os.path.abspath(args.results_dir)
debate_memory_bank: Optional[MemoryBank] = None
if not args.disable_debate_memory and args.debate_memory_dir:
debate_memory_path = Path(args.debate_memory_dir)
if debate_memory_path.exists():
try:
embedding_model = args.embedding_model if args.embedding_model else "BAAI/bge-small-en-v1.5"
debate_memory_bank = MemoryBank(
memory_dir=str(debate_memory_path),
embedding_model=embedding_model
)
except Exception as exc: # noqa: BLE001
print(f"⚠️ Warning: failed to load debate memory from {debate_memory_path}: {exc}")
else:
print(f"ℹ️ Debate memory directory not found: {debate_memory_path} (skipping context retrieval)")
ensure_script(args.debate_script, "Debate script")
ensure_script(args.execute_script, "Execute script")
modelA_runs = collect_runs(args.results_dir, args.modelA)
modelB_runs = collect_runs(args.results_dir, args.modelB)
if args.datasets:
datasets = args.datasets
else:
datasets = sorted(set(modelA_runs.keys()) & set(modelB_runs.keys()))
if not datasets:
print("❌ No common datasets with available runs were found.")
sys.exit(1)
print("=" * 80)
print("🧠 Memory-Based Debate Runner")
print("=" * 80)
print(f"Model A: {args.modelA}")
print(f"Model B: {args.modelB}")
print(f"Datasets: {', '.join(datasets)}")
print(f"Results dir: {args.results_dir}")
print(f"Output root: {args.output_root}")
print(f"Debate workers: {args.debate_workers} (parallel)")
print("=" * 80)
timestamp = dt.datetime.now().strftime("%Y%m%d_%H%M%S")
dataset_summaries: List[Dict] = []
processed = 0
for dataset in datasets:
file_a = pick_latest(modelA_runs, dataset)
file_b = pick_latest(modelB_runs, dataset)
if not file_a or not file_b:
print(f"⚠️ Skipping {dataset}: missing runs for one of the models.")
dataset_summaries.append(
{
"dataset": dataset,
"status": "missing_runs",
"modelA_file": file_a,
"modelB_file": file_b,
}
)
continue
run_dir = os.path.join(
args.output_root,
dataset,
f"{timestamp}_{args.modelA}_vs_{args.modelB}",
)
os.makedirs(run_dir, exist_ok=True)
print(f"\n{'=' * 80}")
print(f"🚀 Dataset: {dataset}")
print(f" Model A file: {file_a}")
print(f" Model B file: {file_b}")
print(f" Output dir: {run_dir}")
print(f"{'=' * 80}")
file_a_for_debate = file_a
file_b_for_debate = file_b
if not args.dry_run and debate_memory_bank and args.debate_memory_top_k > 0:
contexts = build_debate_memory_contexts(
[file_a, file_b], debate_memory_bank, dataset, args.debate_memory_top_k
)
if any(contexts.values()):
print(" 🧠 Injecting debate memory context into prompts")
enriched_a = os.path.join(
run_dir, f"{os.path.basename(file_a)}.debate_memory.jsonl"
)
enriched_b = os.path.join(
run_dir, f"{os.path.basename(file_b)}.debate_memory.jsonl"
)
file_a_for_debate = maybe_enrich_generation_file(file_a, enriched_a, contexts)
file_b_for_debate = maybe_enrich_generation_file(file_b, enriched_b, contexts)
if args.dry_run:
print("Dry-run mode → skipping actual execution.")
dataset_summaries.append(
{
"dataset": dataset,
"status": "dry_run",
"debate_dir": run_dir,
"modelA_file": file_a,
"modelB_file": file_b,
}
)
continue
# 1) Run debate
debate_cmd = [
sys.executable,
"-u",
args.debate_script,
"--resultA",
file_a_for_debate,
"--resultB",
file_b_for_debate,
"--modelA",
args.modelA,
"--modelB",
args.modelB,
"--save_dir",
run_dir,
"--max_rounds",
str(args.max_rounds),
"--temperature",
str(args.temperature),
"--num_workers",
str(args.debate_workers),
]
if args.max_problems is not None:
debate_cmd.extend(["--max_problems", str(args.max_problems)])
debate_log = os.path.join(run_dir, "debate.log")
stream_command(debate_cmd, cwd=str(MONOREPO_ROOT), log_path=debate_log)
consensus_file = os.path.join(
run_dir, f"consensus_{args.modelA}_vs_{args.modelB}.jsonl"
)
if not os.path.exists(consensus_file):
raise FileNotFoundError(
f"Consensus file not found after debate: {consensus_file}"
)
# 2) Evaluate consensus
eval_dir = os.path.join(run_dir, "eval_consensus")
eval_cmd = [
sys.executable,
"-u",
args.execute_script,
"--input_file",
consensus_file,
"--output_dir",
eval_dir,
"--timeout",
str(args.timeout),
"--tolerance",
str(args.tolerance),
"--num_workers",
str(args.execute_workers),
]
if args.relative_tolerance:
eval_cmd.append("--use_relative_tolerance")
if args.save_execution_stdout:
eval_cmd.append("--save_output")
if args.execute_memory_dir:
eval_cmd.extend(["--memory_dir", args.execute_memory_dir])
if args.execute_debug_memory_path:
eval_cmd.extend(["--debug_memory_path", args.execute_debug_memory_path])
if args.execute_disable_debug_memory:
eval_cmd.append("--disable_debug_memory")
if args.embedding_model:
eval_cmd.extend(["--embedding_model", args.embedding_model])
eval_log = os.path.join(run_dir, "evaluate.log")
stream_command(eval_cmd, cwd=str(PROJECT_ROOT), log_path=eval_log)
report_path = os.path.join(eval_dir, "evaluation_report.json")
report = load_eval_report(report_path)
if report is None:
raise FileNotFoundError(f"Missing evaluation report: {report_path}")
dataset_summaries.append(
{
"dataset": dataset,
"status": "completed",
"debate_dir": run_dir,
"accuracy": report.get("accuracy"),
"correct": report.get("correct"),
"total": report.get("total_problems"),
"report_path": report_path,
}
)
processed += 1
print("\n" + "=" * 80)
print("📊 Debate + Evaluation Summary")
print("=" * 80)
for item in dataset_summaries:
dataset = item["dataset"]
status = item["status"]
if status == "completed":
accuracy = item.get("accuracy")
correct = item.get("correct")
total = item.get("total")
print(
f"{dataset:25s} → accuracy {accuracy:.2%} ({correct}/{total}) | dir: {item['debate_dir']}"
)
elif status == "dry_run":
print(f"{dataset:25s} → dry run (planned dir: {item['debate_dir']})")
else:
print(f"{dataset:25s} → {status} (A={item.get('modelA_file')}, B={item.get('modelB_file')})")
print("=" * 80)
if not args.dry_run and processed == 0:
sys.exit("No datasets were processed successfully.")
if __name__ == "__main__":
main()
|