File size: 16,209 Bytes
81d0d00 | 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 | """
Build a mixed continued-pretraining dataset for a code LM.
Sources (streamed from the Hub β no full download):
- UKPLab/SLTrans (LLVM IR <-> source pairs; primary IRCoder signal)
- allenai/peS2o (open scientific text)
- bigcode/the-stack (permissively licensed source code)
Mixing target (token-weighted): 70 / 15 / 15.
Output: JSONL shards under OUT_DIR. Each line:
{"text": "...", "source": "sltrans" | "pes2o" | "the_stack", "meta": {...}}
The token budget is approximate β we use a fast whitespace token estimate by default
to avoid pulling a heavy tokenizer into the streaming loop. Swap in a real tokenizer
(see TOKENIZER section) if you want exact counts against your model's vocab.
Usage:
pip install "datasets>=2.18" huggingface_hub tqdm
huggingface-cli login # SLTrans + the-stack are gated; you must accept their terms
python build_pretrain_dataset.py
"""
from __future__ import annotations
import json
import os
import random
import sys
import time
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import Callable, Iterator
import socket
from datasets import interleave_datasets, load_dataset
from tqdm import tqdm
# Install hf-transfer and set this env var for significantly faster shard downloads.
# pip install hf-transfer
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
# Raise the per-shard download timeout (default 10s is too short for HF CDN under load).
os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "120")
# Apply a 90s socket-level timeout to every connection in this process.
# This covers HF Hub file-listing API calls (which have no timeout by default)
# and prevents indefinite hangs at 'resolving data files'.
socket.setdefaulttimeout(90)
# ============================================================================
# CONFIG
# ============================================================================
@dataclass
class SourceSpec:
name: str # short id used in output records
hf_path: str # HF dataset path
hf_config: str | None # config name, if any
split: str # split to stream
target_fraction: float # share of the total token budget
text_fn: Callable[[dict], str] # extracts the training text from a row
meta_fn: Callable[[dict], dict] # extracts a small metadata dict
# the-stack is organized by language subset; SLTrans by source language.
# If `subsets` is set, we round-robin over them, each loaded as a separate stream.
subsets: list[str] | None = None
# Total tokens in the final dataset (approximate).
TOTAL_TOKEN_BUDGET = 1_500_000_000
# Per-record length filters (in estimated tokens).
MIN_TOKENS_PER_RECORD = 32
MAX_TOKENS_PER_RECORD = 8192
# Output.
OUT_DIR = Path("./pretrain_mix")
SHARD_RECORDS = 50_000 # records per .jsonl shard
SEED = 17
# Reservoir / sampling. We don't reservoir-sample (would require knowing N);
# instead we accept records with probability `keep_prob` per source, tuned so
# the stream yields roughly the target token count before exhaustion. Set to
# 1.0 to take everything until budget is hit.
KEEP_PROB = {
"sltrans": 1.0,
"pes2o": 1.0,
"the_stack": 0.5, # the-stack is huge; subsample to diversify languages
}
# For the-stack, list languages you want represented. Keep this short for a
# focused replication; expand for broader code coverage.
THE_STACK_LANGS = [
"python", "c", "cpp", "rust", "go", "java", "javascript", "typescript",
]
# SLTrans subsets (source languages whose IR we want). None => use default split.
SLTRANS_SUBSETS = [
f"{lang}/{split}"
for lang in ["C", "C++", "D", "Fortran", "Go", "Haskell", "Nim", "Objective-C", "Python", "Rust", "Swift"]
for split in ["Perf_Optimized", "Size_Optimized"]
]
# ============================================================================
# TEXT / META EXTRACTORS
# ============================================================================
# These are intentionally defensive β different dataset versions name fields
# differently. Adjust if a `KeyError` shows up in your run.
def _first_present(row: dict, keys: list[str], default: str = "") -> str:
for k in keys:
if k in row and row[k]:
return row[k]
return default
def sltrans_text(row: dict) -> str:
"""SLTrans pairs source code with its LLVM IR. Concatenate with a marker so
the model learns to associate them β IRCoder-style."""
src = _first_present(row, ["source", "code", "src", "input"])
ir = _first_present(row, ["ir", "llvm_ir", "llvm", "target", "output"])
if not src or not ir:
return ""
return f"<source>\n{src}\n</source>\n<llvm_ir>\n{ir}\n</llvm_ir>"
def sltrans_meta(row: dict) -> dict:
return {
"lang": _first_present(row, ["language", "lang", "source_lang"]),
}
def pes2o_text(row: dict) -> str:
return _first_present(row, ["text", "content"])
def pes2o_meta(row: dict) -> dict:
return {
"id": _first_present(row, ["id", "doc_id"]),
"source": _first_present(row, ["source", "venue"]),
}
def the_stack_text(row: dict) -> str:
return _first_present(row, ["content", "text", "code"])
def the_stack_meta(row: dict) -> dict:
return {
"lang": _first_present(row, ["lang", "language"]),
"repo": _first_present(row, ["max_stars_repo_name", "repo_name"]),
"license": _first_present(row, ["license"]),
}
# ============================================================================
# SOURCES
# ============================================================================
SOURCES: list[SourceSpec] = [
SourceSpec(
name="sltrans",
hf_path="UKPLab/SLTrans",
hf_config=None,
split="train",
target_fraction=0.70,
text_fn=sltrans_text,
meta_fn=sltrans_meta,
subsets=SLTRANS_SUBSETS,
),
SourceSpec(
name="pes2o",
hf_path="allenai/peS2o",
hf_config="v2",
split="train",
target_fraction=0.15,
text_fn=pes2o_text,
meta_fn=pes2o_meta,
),
SourceSpec(
name="the_stack",
hf_path="bigcode/the-stack",
hf_config=None,
# the-stack uses `data_dir` per language rather than HF configs.
split="train",
target_fraction=0.15,
text_fn=the_stack_text,
meta_fn=the_stack_meta,
subsets=THE_STACK_LANGS,
),
]
# ============================================================================
# TOKEN ESTIMATOR
# ============================================================================
# Whitespace-based estimate. For BPE tokenizers, real tokens ~= 1.3x words for
# natural language and ~1.5β2x for code. We bake a 1.5x correction in here so
# the budget is honest enough for planning. If you want exact counts:
#
# from transformers import AutoTokenizer
# tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-3B")
# def estimate_tokens(text: str) -> int:
# return len(tok.encode(text, add_special_tokens=False))
def estimate_tokens(text: str) -> int:
return int(len(text.split()) * 1.5)
# ============================================================================
# STREAMING
# ============================================================================
def open_stream(spec: SourceSpec, subset: str | None):
"""Return an IterableDataset for a (source, subset) pair, or None if unavailable."""
kwargs = {"split": spec.split, "streaming": True}
if spec.hf_config is not None:
kwargs["name"] = spec.hf_config
# the-stack uses data_dir to select a language.
if spec.hf_path == "bigcode/the-stack" and subset is not None:
kwargs["data_dir"] = f"data/{subset}"
# SLTrans subsets are encoded as "Lang/Split" (e.g. "Python/Perf_Optimized").
if spec.hf_path == "UKPLab/SLTrans" and subset is not None:
lang, slt_split = subset.rsplit("/", 1)
kwargs["name"] = lang
kwargs["split"] = slt_split
_TRANSIENT = ("ssl", "timeout", "handshake", "connection", "timed out")
for attempt in range(5):
try:
return load_dataset(spec.hf_path, **kwargs)
except ValueError as e:
if "Bad split" in str(e):
return None
raise
except Exception as e:
if attempt < 4 and any(k in str(e).lower() for k in _TRANSIENT):
time.sleep(2 ** attempt) # 1s, 2s, 4s, 8s
continue
raise
return None
def round_robin(spec: SourceSpec) -> Iterator[dict]:
"""Yield rows from the source, interleaving across subsets if any.
Data-file resolution (the HF Hub HTTP round-trips that show as
'resolving data files') is parallelised across subsets so all
metadata fetches happen concurrently instead of one-by-one.
"""
if not spec.subsets:
ds = open_stream(spec, None)
if ds is not None:
yield from ds
return
# Resolve all subset streams in parallel β resolution is I/O-bound so
# threads eliminate most of the serial 'resolving data files' wait.
with ThreadPoolExecutor(max_workers=min(4, len(spec.subsets))) as pool:
results = list(pool.map(open_stream, [spec] * len(spec.subsets), spec.subsets))
datasets = [ds for ds in results if ds is not None]
if not datasets:
return
yield from interleave_datasets(datasets, stopping_strategy="all_exhausted")
# ============================================================================
# WRITER
# ============================================================================
class ShardWriter:
def __init__(self, out_dir: Path, prefix: str, records_per_shard: int):
self.out_dir = out_dir
self.prefix = prefix
self.records_per_shard = records_per_shard
self.out_dir.mkdir(parents=True, exist_ok=True)
self._shard_idx = 0
self._records_in_shard = 0
self._fh = None
self._open_new_shard()
def _open_new_shard(self) -> None:
if self._fh is not None:
self._fh.close()
path = self.out_dir / f"{self.prefix}-{self._shard_idx:05d}.jsonl"
self._fh = path.open("w", encoding="utf-8")
self._records_in_shard = 0
self._shard_idx += 1
def write(self, record: dict) -> None:
self._fh.write(json.dumps(record, ensure_ascii=False) + "\n")
self._records_in_shard += 1
if self._records_in_shard >= self.records_per_shard:
self._open_new_shard()
def close(self) -> None:
if self._fh is not None:
self._fh.close()
self._fh = None
# ============================================================================
# MAIN
# ============================================================================
def sample_source(spec: SourceSpec, target_tokens: int, writer: ShardWriter,
rng: random.Random) -> int:
"""Stream `spec` until ~target_tokens have been written. Returns tokens written."""
keep_prob = KEEP_PROB.get(spec.name, 1.0)
tokens_written = 0
records_written = 0
rows_seen = 0
rows_skipped_filter = 0
rows_skipped_subsample = 0
rows_skipped_empty = 0
started = time.time()
# tqdm bar measured in tokens β the unit that actually matters for the budget.
# `unit_scale=True` formats large counts as 1.23M / 1.23B automatically.
bar = tqdm(
total=target_tokens,
unit="tok",
unit_scale=True,
desc=f"{spec.name:>10}",
dynamic_ncols=True,
smoothing=0.05,
)
def _refresh_postfix() -> None:
elapsed = max(time.time() - started, 1e-6)
bar.set_postfix({
"records": f"{records_written:,}",
"rows": f"{rows_seen:,}",
"tok/s": f"{tokens_written/elapsed:,.0f}",
"skip": f"{rows_skipped_filter+rows_skipped_subsample+rows_skipped_empty:,}",
}, refresh=False)
try:
for row in round_robin(spec):
rows_seen += 1
if keep_prob < 1.0 and rng.random() > keep_prob:
rows_skipped_subsample += 1
continue
try:
text = spec.text_fn(row)
except Exception as e:
if rows_seen <= 3:
bar.write(f"[{spec.name}] text_fn error on row {rows_seen}: {e!r}")
rows_skipped_empty += 1
continue
if not text:
rows_skipped_empty += 1
continue
n_tok = estimate_tokens(text)
if n_tok < MIN_TOKENS_PER_RECORD or n_tok > MAX_TOKENS_PER_RECORD:
rows_skipped_filter += 1
continue
record = {
"text": text,
"source": spec.name,
"meta": spec.meta_fn(row),
"est_tokens": n_tok,
}
writer.write(record)
tokens_written += n_tok
records_written += 1
# Don't overshoot the bar (tqdm clamps, but `min` keeps the math clean).
bar.update(min(n_tok, target_tokens - bar.n))
# Refresh the postfix every ~1k records β cheaper than every step.
if records_written % 1_000 == 0:
_refresh_postfix()
if tokens_written >= target_tokens:
break
_refresh_postfix()
finally:
bar.close()
elapsed = time.time() - started
print(f"[{spec.name}] DONE: {records_written:,} records, "
f"{tokens_written:,} tokens, {rows_seen:,} rows seen, "
f"skipped(filter={rows_skipped_filter:,} subsample={rows_skipped_subsample:,} "
f"empty={rows_skipped_empty:,}), {elapsed:.0f}s")
return tokens_written
def main() -> None:
rng = random.Random(SEED)
OUT_DIR.mkdir(parents=True, exist_ok=True)
# Sanity-check fractions sum to ~1.
total_frac = sum(s.target_fraction for s in SOURCES)
if abs(total_frac - 1.0) > 1e-3:
print(f"WARN: source fractions sum to {total_frac}, not 1.0", file=sys.stderr)
# Banner β reassures the user something is happening before HF streams open.
print("=" * 64)
print(f"Building mixed pretrain corpus β {OUT_DIR.resolve()}")
print(f"Total token budget: {TOTAL_TOKEN_BUDGET:,}")
print(f"Per-record range: {MIN_TOKENS_PER_RECORD}β{MAX_TOKENS_PER_RECORD} tokens")
for s in SOURCES:
target = int(TOTAL_TOKEN_BUDGET * s.target_fraction)
kp = KEEP_PROB.get(s.name, 1.0)
subs = f", subsets={s.subsets}" if s.subsets else ""
print(f" - {s.name:>10}: {s.target_fraction:>5.0%} β {target:>15,} tok "
f"[keep_prob={kp}{subs}]")
print("=" * 64)
summary = {}
for spec in SOURCES:
target = int(TOTAL_TOKEN_BUDGET * spec.target_fraction)
writer = ShardWriter(OUT_DIR, prefix=spec.name,
records_per_shard=SHARD_RECORDS)
try:
written = sample_source(spec, target, writer, rng)
finally:
writer.close()
summary[spec.name] = {"target": target, "written": written}
# Manifest.
manifest = {
"total_token_budget": TOTAL_TOKEN_BUDGET,
"seed": SEED,
"sources": [
{"name": s.name, "hf_path": s.hf_path,
"fraction": s.target_fraction, "subsets": s.subsets}
for s in SOURCES
],
"tokens_written": summary,
}
(OUT_DIR / "manifest.json").write_text(json.dumps(manifest, indent=2))
print("\n=== SUMMARY ===")
grand_total = sum(v["written"] for v in summary.values())
for name, v in summary.items():
pct = 100 * v["written"] / grand_total if grand_total else 0
print(f" {name:>10}: {v['written']:>15,} tokens ({pct:5.1f}%)")
print(f" {'TOTAL':>10}: {grand_total:>15,} tokens")
print(f"\nOutput: {OUT_DIR.resolve()}")
if __name__ == "__main__":
main() |