File size: 22,047 Bytes
1e59964 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 | """
StepProbe: Targeted Restoration
Given the diagnosis results, construct minimal "Silver Bullet" datasets
and apply QLoRA or DPO fine-tuning to restore reasoning capability
at the identified failure points.
Key insight: We don't need to fine-tune on everything — just on the
specific error patterns that quantization introduces.
"""
import json
import os
import random
from typing import List, Dict, Optional, Tuple
from dataclasses import dataclass, field
from stepprobe.utils import load_jsonl, save_jsonl, set_seed
@dataclass
class RestorationSample:
"""A single training sample for restoration."""
problem_id: str
problem_text: str
# For SFT/QLoRA
correct_cot: str # full correct CoT from FP16
# For DPO
incorrect_cot: Optional[str] = None # incorrect CoT from quantized
# Metadata
error_type: str = ""
failure_step: int = -1
# ============================================================
# Silver Bullet Dataset Construction
# ============================================================
def build_silver_bullet_dataset(
diagnosed_traces: List[dict],
ref_traces: List[dict],
problems: List[dict],
max_samples: int = 500,
target_error_types: Optional[List[str]] = None,
seed: int = 42,
sampling_strategy: str = "silver_bullet",
) -> Tuple[List[RestorationSample], Dict[str, int]]:
"""
Build a training dataset from diagnosed traces.
`sampling_strategy` controls which problems get selected. The three
strategies are used for ablation / baseline comparison so the value of
the step-level diagnosis is falsifiable:
- "silver_bullet" (default): failed problems only, proportional
allocation across the four error types.
- "failed_only": failed problems only, uniform random (no
error-type balancing). Isolates the contribution of error-type
proportional allocation.
- "random": ALL diagnosed problems sampled uniformly (both
failed and correct). Isolates whether diagnosis matters at all
— a negative control.
Args:
diagnosed_traces: Output from diagnose_batch
ref_traces: FP16 segmented traces
problems: Original problems
max_samples: Maximum dataset size
target_error_types: If specified, only include these error types
seed: Random seed
sampling_strategy: One of {"silver_bullet", "failed_only", "random"}
Returns:
(samples, stats) where stats maps error_type -> count
"""
assert sampling_strategy in {"silver_bullet", "failed_only", "random"}, \
f"unknown sampling_strategy: {sampling_strategy}"
set_seed(seed)
ref_by_id = {t["problem_id"]: t for t in ref_traces}
prob_by_id = {p.get("problem_id", p.get("id", "")): p for p in problems}
# Collect candidate samples. For "random", we include correct traces too
# (whose "error type" is tagged as "correct") so the baseline gets to
# sample from the full set.
failures_by_type: Dict[str, List[RestorationSample]] = {
"conceptual": [], "methodological": [], "executional": [], "logical": [],
}
correct_pool: List[RestorationSample] = []
for trace in diagnosed_traces:
pid = trace["problem_id"]
ref = ref_by_id.get(pid)
prob = prob_by_id.get(pid, {})
if ref is None:
continue
is_failure = trace.get("is_correct_final", True) is False
# Identify first error + type (applicable only for failures).
first_error_type = "executional"
first_error_step = -1
for step in trace.get("steps", []):
if step.get("is_correct") is False:
first_error_type = step.get("error_type", "executional")
first_error_step = step.get("index", -1)
break
if target_error_types and is_failure and first_error_type not in target_error_types:
continue
sample = RestorationSample(
problem_id=pid,
problem_text=prob.get("question", prob.get("problem", "")),
correct_cot=ref.get("raw_output", ""),
incorrect_cot=trace.get("raw_output", "") if is_failure else None,
error_type=first_error_type if is_failure else "correct",
failure_step=first_error_step,
)
if is_failure and first_error_type in failures_by_type:
failures_by_type[first_error_type].append(sample)
elif not is_failure:
correct_pool.append(sample)
failed_all = [s for pool in failures_by_type.values() for s in pool]
if sampling_strategy == "random":
pool_all = failed_all + correct_pool
if not pool_all:
print("[WARN] No diagnosed traces available.")
return [], {}
n_alloc = min(max_samples, len(pool_all))
samples = random.sample(pool_all, n_alloc)
stats: Dict[str, int] = {}
for s in samples:
stats[s.error_type] = stats.get(s.error_type, 0) + 1
random.shuffle(samples)
print(f"[random baseline] {len(samples)} samples from {len(pool_all)} diagnosed problems")
for k, v in sorted(stats.items()):
print(f" {k}: {v}")
return samples, stats
if sampling_strategy == "failed_only":
if not failed_all:
print("[WARN] No failures found.")
return [], {}
n_alloc = min(max_samples, len(failed_all))
samples = random.sample(failed_all, n_alloc)
stats = {}
for s in samples:
stats[s.error_type] = stats.get(s.error_type, 0) + 1
random.shuffle(samples)
print(f"[failed-only baseline] {len(samples)} samples from {len(failed_all)} failures")
for k, v in sorted(stats.items()):
print(f" {k}: {v}")
return samples, stats
# Default: silver_bullet (proportional across error types)
total_failures = sum(len(v) for v in failures_by_type.values())
if total_failures == 0:
print("[WARN] No failures found. Nothing to restore.")
return [], {}
samples = []
stats = {}
for etype, pool in failures_by_type.items():
if not pool:
continue
n_alloc = max(1, int(max_samples * len(pool) / total_failures))
n_alloc = min(n_alloc, len(pool))
selected = random.sample(pool, n_alloc)
samples.extend(selected)
stats[etype] = n_alloc
random.shuffle(samples)
samples = samples[:max_samples]
print(f"Silver Bullet dataset: {len(samples)} samples from {total_failures} failures")
for etype, count in sorted(stats.items()):
print(f" {etype}: {count}")
return samples, stats
def format_for_sft(samples: List[RestorationSample]) -> List[dict]:
"""Format samples for supervised fine-tuning (SFT / QLoRA)."""
formatted = []
for s in samples:
formatted.append({
"messages": [
{"role": "user", "content": s.problem_text},
{"role": "assistant", "content": s.correct_cot},
],
"metadata": {
"problem_id": s.problem_id,
"error_type": s.error_type,
"failure_step": s.failure_step,
},
})
return formatted
def format_for_dpo(samples: List[RestorationSample]) -> List[dict]:
"""Format samples for Direct Preference Optimization (DPO)."""
formatted = []
for s in samples:
if not s.incorrect_cot:
continue
formatted.append({
"prompt": s.problem_text,
"chosen": s.correct_cot,
"rejected": s.incorrect_cot,
"metadata": {
"problem_id": s.problem_id,
"error_type": s.error_type,
"failure_step": s.failure_step,
},
})
return formatted
# ============================================================
# QLoRA Fine-Tuning
# ============================================================
def run_qlora_restoration(
model_name: str,
train_data: List[dict],
output_dir: str,
r: int = 16,
lora_alpha: int = 32,
target_modules: List[str] = None,
learning_rate: float = 2e-4,
num_epochs: int = 3,
batch_size: int = 4,
gradient_accumulation_steps: int = 4,
max_seq_length: int = 2048,
):
"""
Run QLoRA fine-tuning for targeted restoration.
This loads the model in 4-bit, applies LoRA adapters,
and fine-tunes on the Silver Bullet dataset.
"""
import torch
from transformers import (
AutoModelForCausalLM, AutoTokenizer,
BitsAndBytesConfig, TrainingArguments,
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer, SFTConfig
from datasets import Dataset
if target_modules is None:
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj"]
os.makedirs(output_dir, exist_ok=True)
print(f"Loading model: {model_name} (4-bit)")
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
device_map="auto",
trust_remote_code=True,
)
model = prepare_model_for_kbit_training(model)
# Apply LoRA
lora_config = LoraConfig(
r=r,
lora_alpha=lora_alpha,
target_modules=target_modules,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Format data
def format_messages(example):
text = tokenizer.apply_chat_template(
example["messages"], tokenize=False, add_generation_prompt=False
)
return {"text": text}
dataset = Dataset.from_list(train_data)
dataset = dataset.map(format_messages)
# Training — save every 25 steps so an OOM crash mid-epoch loses at most
# ~25 steps of work rather than the whole run. save_total_limit=2 keeps
# disk usage bounded (each LoRA checkpoint is only ~30 MB).
training_args = SFTConfig(
output_dir=output_dir,
num_train_epochs=num_epochs,
per_device_train_batch_size=1,
gradient_accumulation_steps=batch_size * gradient_accumulation_steps,
learning_rate=learning_rate,
weight_decay=0.01,
warmup_ratio=0.1,
lr_scheduler_type="cosine",
logging_steps=10,
save_strategy="steps",
save_steps=25,
save_total_limit=2,
bf16=True,
max_length=max_seq_length,
dataset_text_field="text",
gradient_checkpointing=True,
report_to="none",
)
trainer = SFTTrainer(
model=model,
processing_class=tokenizer,
train_dataset=dataset,
args=training_args,
)
# Auto-resume from the latest checkpoint if one exists. HuggingFace's
# `resume_from_checkpoint=True` fails loudly when no checkpoint is on
# disk, so we detect by hand.
import glob as _glob
existing_ckpts = sorted(
_glob.glob(os.path.join(output_dir, "checkpoint-*")),
key=lambda p: int(p.rsplit("-", 1)[-1]) if p.rsplit("-", 1)[-1].isdigit() else -1,
)
if existing_ckpts:
latest = existing_ckpts[-1]
print(f"Resuming from checkpoint: {latest}")
print(f"Starting QLoRA training: {len(train_data)} samples, {num_epochs} epochs (resume)")
trainer.train(resume_from_checkpoint=latest)
else:
print(f"Starting QLoRA training: {len(train_data)} samples, {num_epochs} epochs")
trainer.train()
# Save final adapter at a stable path the rest of the pipeline looks for.
adapter_path = os.path.join(output_dir, "adapter")
model.save_pretrained(adapter_path)
tokenizer.save_pretrained(adapter_path)
print(f"Adapter saved to {adapter_path}")
# Clean up intermediate checkpoints so subsequent runs don't accidentally
# resume stale training on top of a newer dataset (and to save disk).
for ck in _glob.glob(os.path.join(output_dir, "checkpoint-*")):
import shutil
shutil.rmtree(ck, ignore_errors=True)
return adapter_path
# ============================================================
# DPO Fine-Tuning
# ============================================================
def run_dpo_restoration(
model_name: str,
train_data: List[dict],
output_dir: str,
beta: float = 0.1,
learning_rate: float = 5e-5,
num_epochs: int = 1,
batch_size: int = 2,
gradient_accumulation_steps: int = 8,
max_seq_length: int = 1024,
):
"""
Run DPO fine-tuning using (correct, incorrect) CoT pairs.
"""
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, prepare_model_for_kbit_training
from trl import DPOTrainer, DPOConfig
from datasets import Dataset
os.makedirs(output_dir, exist_ok=True)
print(f"Loading model: {model_name} (4-bit)")
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
device_map="auto",
trust_remote_code=True,
)
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
)
# Build DPO dataset
dataset = Dataset.from_list(train_data)
training_args = DPOConfig(
output_dir=output_dir,
num_train_epochs=num_epochs,
per_device_train_batch_size=1,
gradient_accumulation_steps=batch_size * gradient_accumulation_steps,
learning_rate=learning_rate,
beta=beta,
warmup_ratio=0.1,
lr_scheduler_type="cosine",
logging_steps=10,
save_strategy="steps",
save_steps=25,
save_total_limit=2,
bf16=True,
max_length=max_seq_length,
gradient_checkpointing=True,
report_to="none",
)
trainer = DPOTrainer(
model=model,
ref_model=None, # use implicit reference with peft
processing_class=tokenizer,
train_dataset=dataset,
args=training_args,
peft_config=lora_config,
)
# Auto-resume from the latest checkpoint if one exists.
import glob as _glob
existing_ckpts = sorted(
_glob.glob(os.path.join(output_dir, "checkpoint-*")),
key=lambda p: int(p.rsplit("-", 1)[-1]) if p.rsplit("-", 1)[-1].isdigit() else -1,
)
if existing_ckpts:
latest = existing_ckpts[-1]
print(f"Resuming DPO training from checkpoint: {latest}")
trainer.train(resume_from_checkpoint=latest)
else:
print(f"Starting DPO training: {len(train_data)} pairs, beta={beta}")
trainer.train()
adapter_path = os.path.join(output_dir, "adapter")
trainer.save_model(adapter_path)
tokenizer.save_pretrained(adapter_path)
print(f"Adapter saved to {adapter_path}")
# Drop intermediate checkpoints once the final adapter is on disk.
for ck in _glob.glob(os.path.join(output_dir, "checkpoint-*")):
import shutil
shutil.rmtree(ck, ignore_errors=True)
return adapter_path
# ============================================================
# CLI
# ============================================================
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Targeted restoration of quantized reasoning models")
parser.add_argument("--model", required=True, help="Base model name")
parser.add_argument("--diagnosis", required=True, help="Directory with diagnosed traces")
parser.add_argument("--ref", required=True, help="Directory with FP16 segmented traces")
parser.add_argument("--problems", default=None, help="JSONL with original problems")
parser.add_argument("--output", required=True, help="Output directory")
parser.add_argument("--method", default="qlora", choices=["qlora", "dpo"])
parser.add_argument("--max-samples", type=int, default=500)
parser.add_argument("--target-errors", nargs="*", default=None,
help="Target specific error types (e.g., executional conceptual)")
parser.add_argument("--sampling-strategy", default="silver_bullet",
choices=["silver_bullet", "failed_only", "random"],
help="Dataset construction: silver_bullet (failures "
"balanced across error types), failed_only (failures "
"uniformly sampled), or random (sample from ALL "
"diagnosed problems including correct ones). The "
"latter two are baselines against which silver_bullet "
"is measured.")
parser.add_argument("--epochs", type=int, default=3)
parser.add_argument("--lr", type=float, default=2e-4)
parser.add_argument("--batch-size", type=int, default=4)
parser.add_argument("--max-seq-length", type=int, default=2048,
help="Max sequence length during training. Drop to 1024 "
"if training OOMs under memory pressure.")
parser.add_argument("--lora-rank", type=int, default=16,
help="LoRA rank. Halve (8) if training OOMs.")
args = parser.parse_args()
import glob
# Load data
diagnosed = []
for f in sorted(glob.glob(os.path.join(args.diagnosis, "*.jsonl"))):
diagnosed.extend(load_jsonl(f))
ref_traces = []
for f in sorted(glob.glob(os.path.join(args.ref, "*.jsonl"))):
ref_traces.extend(load_jsonl(f))
problems = []
if args.problems:
problems = load_jsonl(args.problems)
else:
# Reconstruct questions from the original benchmark datasets.
# problem_ids are like "gsm8k_0", "math500_123", "gpqa_42".
from datasets import load_dataset
_bench_cache = {}
def _load_bench(bench_name):
if bench_name in _bench_cache:
return _bench_cache[bench_name]
qs = []
if bench_name == "gsm8k":
ds = load_dataset("openai/gsm8k", "main", split="test")
qs = [ex["question"] for ex in ds]
elif bench_name == "math500":
ds = load_dataset("HuggingFaceH4/MATH-500", split="test")
qs = [ex["problem"] for ex in ds]
elif bench_name == "gpqa":
ds = load_dataset("Idavidrein/gpqa", "gpqa_diamond", split="train")
qs = [ex["Question"] for ex in ds]
_bench_cache[bench_name] = qs
return qs
def _get_question(problem_id: str) -> str:
parts = problem_id.rsplit("_", 1)
if len(parts) != 2 or not parts[1].isdigit():
return ""
bench_name, idx = parts[0], int(parts[1])
qs = _load_bench(bench_name)
return qs[idx] if idx < len(qs) else ""
for t in ref_traces:
pid = t["problem_id"]
problems.append({
"problem_id": pid,
"question": _get_question(pid),
"answer": t.get("final_answer", ""),
})
# Build training dataset under the chosen sampling strategy.
samples, stats = build_silver_bullet_dataset(
diagnosed_traces=diagnosed,
ref_traces=ref_traces,
sampling_strategy=args.sampling_strategy,
problems=problems,
max_samples=args.max_samples,
target_error_types=args.target_errors,
)
if not samples:
print("No samples to train on. Exiting.")
exit(0)
# Save dataset
dataset_dir = os.path.join(args.output, "dataset")
os.makedirs(dataset_dir, exist_ok=True)
if args.method == "qlora":
train_data = format_for_sft(samples)
save_jsonl(train_data, os.path.join(dataset_dir, "sft_train.jsonl"))
print(f"\nSFT dataset saved: {len(train_data)} samples")
adapter_path = run_qlora_restoration(
model_name=args.model,
train_data=train_data,
output_dir=os.path.join(args.output, "qlora"),
learning_rate=args.lr,
num_epochs=args.epochs,
batch_size=args.batch_size,
r=args.lora_rank,
lora_alpha=args.lora_rank * 2,
max_seq_length=args.max_seq_length,
)
elif args.method == "dpo":
train_data = format_for_dpo(samples)
save_jsonl(train_data, os.path.join(dataset_dir, "dpo_train.jsonl"))
print(f"\nDPO dataset saved: {len(train_data)} pairs")
adapter_path = run_dpo_restoration(
model_name=args.model,
train_data=train_data,
output_dir=os.path.join(args.output, "dpo"),
learning_rate=args.lr,
num_epochs=args.epochs,
batch_size=args.batch_size,
)
# Save stats
save_jsonl([{"stats": stats, "n_samples": len(samples), "method": args.method}],
os.path.join(args.output, "restoration_stats.jsonl"))
print(f"\nRestoration complete! Adapter at: {adapter_path}")
|