Spaces:
Paused
Paused
File size: 16,436 Bytes
45ee481 | 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 | """
Dataset Builder Module
Build final training dataset in ChatML format for Qwen3 fine-tuning.
Creates train/validation splits with proper formatting.
Example usage:
builder = DatasetBuilder(system_prompt="You are Ryouken Okuni...")
builder.build_from_qa_pairs(qa_pairs, output_dir="data/training/")
"""
import json
import random
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
from loguru import logger
try:
import tiktoken
TIKTOKEN_AVAILABLE = True
except ImportError:
TIKTOKEN_AVAILABLE = False
@dataclass
class DatasetStatistics:
"""Statistics about the built dataset."""
total_examples: int
train_examples: int
validation_examples: int
avg_tokens_per_example: float
max_tokens: int
min_tokens: int
total_tokens: int
question_type_distribution: dict
def to_dict(self) -> dict:
"""Convert to dictionary for serialization."""
return {
"total_examples": self.total_examples,
"train_examples": self.train_examples,
"validation_examples": self.validation_examples,
"avg_tokens_per_example": round(self.avg_tokens_per_example, 2),
"max_tokens": self.max_tokens,
"min_tokens": self.min_tokens,
"total_tokens": self.total_tokens,
"question_type_distribution": self.question_type_distribution,
}
class DatasetBuilder:
"""
Build training datasets in ChatML format for Qwen3.
Features:
- ChatML message format
- Train/validation split
- Deduplication
- Token count validation
- Statistics generation
Example:
>>> builder = DatasetBuilder()
>>> stats = builder.build_from_qa_pairs(qa_pairs, "data/training/")
>>> print(f"Built {stats.total_examples} examples")
"""
# Default system prompt template
DEFAULT_SYSTEM_PROMPT = """You are {ceo_name}, CEO of {company_name}.
You are a visionary technology leader with deep expertise in AI, business strategy, and innovation. Your communication style is thoughtful, confident, and grounded in real-world experience.
Key traits:
- You explain complex concepts clearly using analogies and examples
- You balance strategic thinking with practical insights
- You are passionate about technology's potential to transform business
- You value authenticity and speak from genuine experience
- You are direct but respectful in your communication
When responding:
- Draw from your extensive experience in technology and business
- Share insights that reflect your unique perspective as a CEO
- Be helpful and substantive in your answers
- Maintain a professional yet personable tone appropriate for Japanese business culture"""
def __init__(
self,
system_prompt: Optional[str] = None,
ceo_name: str = "Ryouken Okuni",
company_name: str = "Akatsuki AI Technologies",
max_tokens_per_example: int = 2048,
encoding_name: str = "cl100k_base",
):
"""
Initialize the dataset builder.
Args:
system_prompt: Custom system prompt (uses default if None)
ceo_name: CEO name to insert into prompt
company_name: Company name to insert into prompt
max_tokens_per_example: Maximum tokens per training example
encoding_name: Tiktoken encoding name
"""
self.ceo_name = ceo_name
self.company_name = company_name
self.max_tokens_per_example = max_tokens_per_example
# Set system prompt
if system_prompt:
self.system_prompt = system_prompt
else:
self.system_prompt = self.DEFAULT_SYSTEM_PROMPT.format(
ceo_name=ceo_name,
company_name=company_name,
)
# Initialize tokenizer
if TIKTOKEN_AVAILABLE:
try:
self.encoding = tiktoken.get_encoding(encoding_name)
except Exception:
self.encoding = None
else:
self.encoding = None
def count_tokens(self, text: str) -> int:
"""Count tokens in text."""
if self.encoding:
return len(self.encoding.encode(text))
return len(text) // 3 # Rough approximation
def build_from_qa_pairs(
self,
qa_pairs: list,
output_dir: str | Path,
train_ratio: float = 0.9,
shuffle: bool = True,
deduplicate: bool = True,
) -> DatasetStatistics:
"""
Build training dataset from Q&A pairs.
Args:
qa_pairs: List of QAPair objects or dicts
output_dir: Directory to save train.jsonl and validation.jsonl
train_ratio: Ratio for train/validation split (default 0.9)
shuffle: Whether to shuffle data
deduplicate: Whether to remove duplicate questions
Returns:
DatasetStatistics object
"""
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
logger.info(f"Building dataset from {len(qa_pairs)} Q&A pairs")
# Convert to standard format
examples = self._convert_qa_pairs(qa_pairs)
# Deduplicate
if deduplicate:
original_count = len(examples)
examples = self._deduplicate(examples)
logger.info(f"Deduplication: {original_count} -> {len(examples)} examples")
# Validate token counts
examples = self._validate_token_counts(examples)
logger.info(f"After token validation: {len(examples)} examples")
# Shuffle
if shuffle:
random.shuffle(examples)
# Split into train/validation
split_idx = int(len(examples) * train_ratio)
train_examples = examples[:split_idx]
val_examples = examples[split_idx:]
# Save datasets
train_path = output_dir / "train.jsonl"
val_path = output_dir / "validation.jsonl"
self._save_jsonl(train_examples, train_path)
self._save_jsonl(val_examples, val_path)
# Calculate statistics
stats = self._calculate_statistics(examples, train_examples, val_examples)
# Save statistics
stats_path = output_dir / "dataset_stats.json"
with open(stats_path, "w", encoding="utf-8") as f:
json.dump(stats.to_dict(), f, indent=2)
logger.info(f"Saved train set: {train_path} ({len(train_examples)} examples)")
logger.info(f"Saved validation set: {val_path} ({len(val_examples)} examples)")
logger.info(f"Saved statistics: {stats_path}")
return stats
def _convert_qa_pairs(self, qa_pairs: list) -> list[dict]:
"""Convert Q&A pairs to ChatML format."""
examples = []
for pair in qa_pairs:
# Handle both QAPair objects and dicts
if hasattr(pair, "question"):
question = pair.question
answer = pair.answer
q_type = pair.question_type
else:
question = pair["question"]
answer = pair["answer"]
q_type = pair.get("question_type", "unknown")
example = {
"messages": [
{"role": "system", "content": self.system_prompt},
{"role": "user", "content": question},
{"role": "assistant", "content": answer},
],
"metadata": {
"question_type": q_type,
},
}
examples.append(example)
return examples
def _deduplicate(self, examples: list[dict]) -> list[dict]:
"""Remove examples with duplicate questions."""
seen_questions = set()
unique_examples = []
for example in examples:
# Get user message (the question)
question = None
for msg in example["messages"]:
if msg["role"] == "user":
question = msg["content"].strip().lower()
break
if question and question not in seen_questions:
seen_questions.add(question)
unique_examples.append(example)
return unique_examples
def _validate_token_counts(self, examples: list[dict]) -> list[dict]:
"""Filter out examples that exceed token limit."""
valid_examples = []
for example in examples:
# Calculate total tokens
total_tokens = 0
for msg in example["messages"]:
total_tokens += self.count_tokens(msg["content"])
total_tokens += 4 # Approximate overhead per message
if total_tokens <= self.max_tokens_per_example:
example["token_count"] = total_tokens
valid_examples.append(example)
else:
logger.debug(f"Skipping example with {total_tokens} tokens (max: {self.max_tokens_per_example})")
return valid_examples
def _save_jsonl(self, examples: list[dict], path: Path) -> None:
"""Save examples to JSONL format."""
with open(path, "w", encoding="utf-8") as f:
for example in examples:
# Remove metadata before saving (keep only messages)
output = {"messages": example["messages"]}
f.write(json.dumps(output, ensure_ascii=False) + "\n")
def _calculate_statistics(
self,
all_examples: list[dict],
train_examples: list[dict],
val_examples: list[dict],
) -> DatasetStatistics:
"""Calculate dataset statistics."""
token_counts = [ex.get("token_count", 0) for ex in all_examples]
# Question type distribution
type_counts = {}
for ex in all_examples:
q_type = ex.get("metadata", {}).get("question_type", "unknown")
type_counts[q_type] = type_counts.get(q_type, 0) + 1
return DatasetStatistics(
total_examples=len(all_examples),
train_examples=len(train_examples),
validation_examples=len(val_examples),
avg_tokens_per_example=sum(token_counts) / len(token_counts) if token_counts else 0,
max_tokens=max(token_counts) if token_counts else 0,
min_tokens=min(token_counts) if token_counts else 0,
total_tokens=sum(token_counts),
question_type_distribution=type_counts,
)
def build_from_segments(
self,
segments: list,
output_dir: str | Path,
train_ratio: float = 0.9,
) -> DatasetStatistics:
"""
Build training dataset directly from text segments (for continuation training).
This creates examples where the model learns to continue CEO-style text.
Args:
segments: List of TextSegment objects or dicts
output_dir: Directory to save datasets
train_ratio: Train/validation split ratio
Returns:
DatasetStatistics object
"""
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
logger.info(f"Building continuation dataset from {len(segments)} segments")
examples = []
for segment in segments:
content = segment.content if hasattr(segment, "content") else segment["content"]
# Create a simple prompt asking to continue the thought
example = {
"messages": [
{"role": "system", "content": self.system_prompt},
{"role": "user", "content": "Please share your thoughts on this topic."},
{"role": "assistant", "content": content},
],
"metadata": {"type": "continuation"},
}
examples.append(example)
# Validate and save
examples = self._validate_token_counts(examples)
random.shuffle(examples)
split_idx = int(len(examples) * train_ratio)
train_examples = examples[:split_idx]
val_examples = examples[split_idx:]
self._save_jsonl(train_examples, output_dir / "train.jsonl")
self._save_jsonl(val_examples, output_dir / "validation.jsonl")
stats = self._calculate_statistics(examples, train_examples, val_examples)
with open(output_dir / "dataset_stats.json", "w", encoding="utf-8") as f:
json.dump(stats.to_dict(), f, indent=2)
return stats
@staticmethod
def load_dataset(path: str | Path) -> list[dict]:
"""Load a JSONL dataset file."""
examples = []
with open(path, "r", encoding="utf-8") as f:
for line in f:
if line.strip():
examples.append(json.loads(line))
return examples
def update_system_prompt(self, new_prompt: str) -> None:
"""Update the system prompt for future builds."""
self.system_prompt = new_prompt
logger.info("System prompt updated")
def get_system_prompt(self) -> str:
"""Get the current system prompt."""
return self.system_prompt
def main():
"""CLI entry point for testing the builder."""
import argparse
parser = argparse.ArgumentParser(
description="Build training datasets in ChatML format",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
python dataset_builder.py qa_pairs.json --output data/training/
python dataset_builder.py qa_pairs.json --train-ratio 0.85
python dataset_builder.py qa_pairs.json --system-prompt "Custom prompt..."
Input format (qa_pairs.json):
[
{"question": "...", "answer": "...", "question_type": "..."},
...
]
Output format (train.jsonl):
{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
""",
)
parser.add_argument("input", help="Input Q&A pairs JSON file")
parser.add_argument(
"--output",
"-o",
default="data/training/",
help="Output directory (default: data/training/)",
)
parser.add_argument(
"--train-ratio",
type=float,
default=0.9,
help="Train/validation split ratio (default: 0.9)",
)
parser.add_argument(
"--system-prompt",
help="Custom system prompt (uses default if not provided)",
)
parser.add_argument(
"--ceo-name",
default="Ryouken Okuni",
help="CEO name for default prompt",
)
parser.add_argument(
"--company-name",
default="Akatsuki AI Technologies",
help="Company name for default prompt",
)
parser.add_argument(
"--max-tokens",
type=int,
default=2048,
help="Maximum tokens per example (default: 2048)",
)
parser.add_argument(
"--no-shuffle",
action="store_true",
help="Don't shuffle the data",
)
parser.add_argument(
"--no-dedup",
action="store_true",
help="Don't deduplicate questions",
)
args = parser.parse_args()
# Load Q&A pairs
with open(args.input, "r", encoding="utf-8") as f:
qa_pairs = json.load(f)
print(f"Loaded {len(qa_pairs)} Q&A pairs")
# Build dataset
builder = DatasetBuilder(
system_prompt=args.system_prompt,
ceo_name=args.ceo_name,
company_name=args.company_name,
max_tokens_per_example=args.max_tokens,
)
stats = builder.build_from_qa_pairs(
qa_pairs=qa_pairs,
output_dir=args.output,
train_ratio=args.train_ratio,
shuffle=not args.no_shuffle,
deduplicate=not args.no_dedup,
)
# Print statistics
print("\n=== Dataset Statistics ===")
print(f"Total examples: {stats.total_examples}")
print(f"Train examples: {stats.train_examples}")
print(f"Validation examples: {stats.validation_examples}")
print(f"Avg tokens/example: {stats.avg_tokens_per_example:.1f}")
print(f"Token range: {stats.min_tokens} - {stats.max_tokens}")
print(f"Total tokens: {stats.total_tokens:,}")
print("\nQuestion type distribution:")
for q_type, count in stats.question_type_distribution.items():
print(f" {q_type}: {count}")
if __name__ == "__main__":
main()
|