Spaces:
Running
Running
File size: 15,150 Bytes
5dbca28 | 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 | """
QA Model Training Script
=========================
Fine-tunes a QA model (roberta-base or squad2 warm-start) on the CUAD dataset.
CUAD contexts are full contracts (avg 54K chars).
We use sliding window tokenization to create 384-token windows
with 128-token overlap β this is how SQuAD-style models handle long documents.
Memory Safety:
- Default batch_size=2 to avoid RAM exhaustion on laptops
- Default max_train_samples=500 (use --max_train_samples=-1 for full dataset)
- Aggressive garbage collection after data transformations
- Tokenization uses small batch sizes to limit peak memory
Usage:
python -m src.train_qa # defaults (safe)
python -m src.train_qa --epochs 3 --batch_size 4
python -m src.train_qa --base_model deepset/roberta-base-squad2
python -m src.train_qa --max_train_samples -1 # full dataset (needs 16GB+ RAM)
"""
import argparse
import functools
import gc
import json
import logging
import os
import sys
from typing import Any, Dict, List, Tuple
logger = logging.getLogger(__name__)
def load_cuad_data(filepath: str) -> Dict:
"""Load a CUAD JSON file (SQuAD 2.0 format)."""
with open(filepath, "r", encoding="utf-8") as f:
data = json.load(f)
return data
def cuad_to_squad_examples(data: Dict) -> List[Dict]:
"""Convert CUAD data to a flat list of SQuAD-style examples.
Each example has: id, question, context, answers, is_impossible.
"""
examples = []
for article in data["data"]:
title = article.get("title", "")
for paragraph in article["paragraphs"]:
context = paragraph["context"]
for qa in paragraph["qas"]:
example = {
"id": qa["id"],
"title": title,
"question": qa["question"],
"context": context,
"answers": {
"text": [a["text"] for a in qa.get("answers", [])],
"answer_start": [a["answer_start"] for a in qa.get("answers", [])],
},
"is_impossible": qa.get("is_impossible", False),
}
examples.append(example)
return examples
def prepare_train_features(examples, tokenizer, max_length=384, doc_stride=128):
"""Prepare training features with sliding window tokenization.
Handles long documents by creating overlapping windows.
Maps answer spans to the correct window positions.
"""
pad_on_right = tokenizer.padding_side == "right"
tokenized = tokenizer(
examples["question"] if pad_on_right else examples["context"],
examples["context"] if pad_on_right else examples["question"],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_length,
stride=doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length",
)
sample_mapping = tokenized.pop("overflow_to_sample_mapping")
offset_mapping = tokenized.pop("offset_mapping")
tokenized["start_positions"] = []
tokenized["end_positions"] = []
for i, offsets in enumerate(offset_mapping):
input_ids = tokenized["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
sequence_ids = tokenized.sequence_ids(i)
sample_index = sample_mapping[i]
answers = examples["answers"][sample_index]
if examples["is_impossible"][sample_index] or len(answers["answer_start"]) == 0:
tokenized["start_positions"].append(cls_index)
tokenized["end_positions"].append(cls_index)
else:
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
token_start_index = 0
while sequence_ids[token_start_index] != (1 if pad_on_right else 0):
token_start_index += 1
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != (1 if pad_on_right else 0):
token_end_index -= 1
if not (
offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char
):
tokenized["start_positions"].append(cls_index)
tokenized["end_positions"].append(cls_index)
else:
while (
token_start_index < len(offsets)
and offsets[token_start_index][0] <= start_char
):
token_start_index += 1
tokenized["start_positions"].append(token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized["end_positions"].append(token_end_index + 1)
return tokenized
def _tokenize_wrapper(examples, tokenizer, max_length, doc_stride):
"""Top-level wrapper so it can be pickled for multiprocessing on Windows."""
return prepare_train_features(examples, tokenizer, max_length, doc_stride)
def train(
train_path: str = "data/train.json",
test_path: str = "data/test.json",
base_model: str = "deepset/roberta-base-squad2",
output_dir: str = "ckpt_obligation",
epochs: int = 2,
batch_size: int = 16,
learning_rate: float = 3e-5,
max_length: int = 384,
doc_stride: int = 128,
max_train_samples: int = None,
device: str = "auto",
):
"""Fine-tune a QA model on CUAD data.
Args:
train_path: Path to CUAD train.json
test_path: Path to CUAD test.json
base_model: HuggingFace model name to fine-tune
output_dir: Directory to save the fine-tuned model
epochs: Number of training epochs (2 is enough when warm-starting from squad2)
batch_size: Training batch size (16 fits in RTX 4050 6GB VRAM with fp16)
learning_rate: Learning rate
max_length: Maximum sequence length for tokenizer
doc_stride: Sliding window stride for long documents
max_train_samples: Limit training samples (default 500, use -1 for all)
device: 'auto' (detect GPU), 'cuda', or 'cpu'
"""
# βββ Imports (done here to allow module to be imported without torch) β
try:
import torch
from datasets import Dataset
from transformers import (
AutoModelForQuestionAnswering,
AutoTokenizer,
TrainingArguments,
Trainer,
default_data_collator,
)
except ImportError as e:
print(f"Missing dependency: {e}")
print("Install with: pip install torch transformers datasets")
sys.exit(1)
# βββ Resolve device ββββββββββββββββββββββββββββββββββββββββββββββββββ
from all_model_code.model_1_code.utils import get_device, get_safe_train_samples
device = get_device(device)
logger.info(f"Training device: {device}")
# βββ Handle max_train_samples βββββββββββββββββββββββββββββββββββββββββ
# -1 means "force all regardless of RAM" (user knows what they're doing)
force_all = (max_train_samples is not None and max_train_samples < 0)
if force_all:
max_train_samples = None
# βββ Load Data βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
logger.info(f"Loading training data from {train_path}")
train_data = load_cuad_data(train_path)
train_examples = cuad_to_squad_examples(train_data)
logger.info(f" β {len(train_examples)} training examples")
# Free the raw JSON immediately β it's huge and no longer needed
del train_data
gc.collect()
# NOTE: test data is NOT loaded here to save memory.
# Evaluation should be done separately via src.evaluate.
# βββ Auto-detect safe sample count if no explicit limit ββββββββββββββ
if max_train_samples is None and not force_all:
max_train_samples = get_safe_train_samples(len(train_examples))
# βββ Create HuggingFace Datasets βββββββββββββββββββββββββββββββββββββ
# Convert to column-oriented format for HF datasets
def examples_to_columns(examples):
columns = {
"id": [], "question": [], "context": [],
"answers": [], "is_impossible": [],
}
for ex in examples:
columns["id"].append(ex["id"])
columns["question"].append(ex["question"])
columns["context"].append(ex["context"])
columns["answers"].append(ex["answers"])
columns["is_impossible"].append(ex["is_impossible"])
return columns
train_dataset = Dataset.from_dict(examples_to_columns(train_examples))
# Free the examples list β dataset holds the data now
del train_examples
gc.collect()
if max_train_samples and max_train_samples < len(train_dataset):
train_dataset = train_dataset.select(range(max_train_samples))
logger.info(f" Using {len(train_dataset)} training samples")
# βββ Load Tokenizer & Model ββββββββββββββββββββββββββββββββββββββββββ
logger.info(f"Loading model: {base_model}")
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForQuestionAnswering.from_pretrained(base_model)
# βββ Tokenize (with disk cache) ββββββββββββββββββββββββββββββββββββ
# CUAD contracts are huge (~54K chars each Γ 22K examples).
# Tokenization takes ~50 min, so we cache to disk after the first run.
# Cache is keyed by sample count so changing --max_train_samples auto-invalidates.
os.environ["TOKENIZERS_PARALLELISM"] = "true" # Rust-level threading
num_samples = len(train_dataset)
cache_dir = os.path.join(output_dir, "tokenized_cache")
cache_path = os.path.join(cache_dir, f"tokenized_train_{num_samples}")
if os.path.exists(cache_path):
from datasets import load_from_disk
logger.info(f"Loading cached tokenized data from {cache_path}")
tokenized_train = load_from_disk(cache_path)
logger.info(f" β {len(tokenized_train)} cached features loaded instantly!")
else:
logger.info(f"Tokenizing {num_samples} training examples (sliding window) β this only happens once per sample count...")
tokenized_train = train_dataset.map(
lambda ex: prepare_train_features(ex, tokenizer, max_length, doc_stride),
batched=True,
batch_size=100, # small batch to limit peak memory
remove_columns=train_dataset.column_names,
desc="Tokenizing",
)
# Save to disk so next run skips this entirely
os.makedirs(cache_dir, exist_ok=True)
tokenized_train.save_to_disk(cache_path)
logger.info(f" β Tokenized data cached to {cache_path}")
# Free the un-tokenized dataset
del train_dataset
gc.collect()
logger.info(f" β {len(tokenized_train)} tokenized features (from sliding windows)")
# βββ Training Arguments ββββββββββββββββββββββββββββββββββββββββββββββ
use_gpu = (device == "cuda")
training_args = TrainingArguments(
output_dir=output_dir,
overwrite_output_dir=True,
num_train_epochs=epochs,
per_device_train_batch_size=batch_size,
learning_rate=learning_rate,
weight_decay=0.01,
warmup_ratio=0.1,
logging_steps=50,
save_strategy="epoch",
save_total_limit=1, # keep only 1 checkpoint to save disk/memory
fp16=use_gpu, # FP16 on GPU for ~2x memory savings
report_to="none",
use_cpu=(not use_gpu),
dataloader_num_workers=0, # avoid multiprocessing memory overhead on Windows
dataloader_pin_memory=use_gpu, # pin_memory speeds up GPU, wastes RAM on CPU
)
# βββ Trainer βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_train,
tokenizer=tokenizer,
data_collator=default_data_collator,
)
# βββ Train βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
logger.info("Starting training...")
trainer.train()
# βββ Save ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
logger.info(f"Saving model to {output_dir}")
trainer.save_model(output_dir)
tokenizer.save_pretrained(output_dir)
logger.info("Training complete!")
return output_dir
def main():
parser = argparse.ArgumentParser(description="Fine-tune QA model on CUAD")
parser.add_argument("--train_path", default="data/train.json")
parser.add_argument("--test_path", default="data/test.json")
parser.add_argument("--base_model", default="deepset/roberta-base-squad2")
parser.add_argument("--output_dir", default="ckpt_obligation")
parser.add_argument("--epochs", type=int, default=2,
help="Training epochs (2 is enough when warm-starting from squad2)")
parser.add_argument("--batch_size", type=int, default=16,
help="Batch size (16 fits RTX 4050 6GB VRAM with fp16)")
parser.add_argument("--learning_rate", type=float, default=3e-5)
parser.add_argument("--max_length", type=int, default=384)
parser.add_argument("--doc_stride", type=int, default=128)
parser.add_argument("--max_train_samples", type=int, default=None,
help="Limit training samples. Default: auto-detect based on RAM. Use -1 to force ALL.")
parser.add_argument("--device", default="auto", choices=["auto", "cpu", "cuda"],
help="Device: 'auto' detects GPU, 'cuda' forces GPU, 'cpu' forces CPU")
args = parser.parse_args()
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
)
train(**vars(args))
if __name__ == "__main__":
main()
|