"""Corpus preparation: stream, decontaminate, tokenize, shard. python prepare.py plan # token targets per source python prepare.py sample --mb 2000 # text sample for the tokenizer python prepare.py source --name ultra_fineweb --tokens 120e9 python prepare.py val --tokens 20e6 python prepare.py status Sources are streamed from the Hub and written straight out as uint16 tokens. Nothing keeps the raw text: 200B tokens is ~400GB stored, and the text it came from is more than twice that -- there is 664GB free on the Media SSD and 20GB on /home, so the token files live on Media and the text never lands at all. Each source becomes data//tokens.bin (uint16) plus docs.idx (uint64 document start offsets). Both are appended to incrementally and a progress file records how far the stream got, so preparation is resumable and training can start on the first sources while the rest are still arriving. Decontamination runs inline: any document sharing a 13-gram with an evaluation item is dropped before it is ever written. Filtering afterwards would mean rewriting a 200GB file. """ from __future__ import annotations import argparse import hashlib import json import re import os import sys import time from pathlib import Path import numpy as np from config import DATA_DIR, Config from tokenizer import BorisTokenizer os.environ.setdefault("TOKENIZERS_PARALLELISM", "true") # name -> (repo, configs, split, text field). Resolved against the Hub on # 2026-09-06 with probe/sources_probe.py, not from memory -- Ultra-FineWeb # splits by language rather than "train", and finephrase has five configs, so # both of those failed when they were guessed. Run `prepare.py check` after any # edit here; it opens every source and pulls one record. # # `configs` may be a tuple: the source is consumed in that order until its token # target is met. Cosmopedia needs this -- no single config holds enough. SOURCES = { "ultra_fineweb": ("openbmb/Ultra-FineWeb", ("default",), "en", "content"), "cosmopedia": ("HuggingFaceTB/cosmopedia", ("web_samples_v2", "web_samples_v1", "auto_math_text", "stories", "openstax", "stanford", "wikihow", "khanacademy"), "train", "text"), "finephrase": ("HuggingFaceFW/finephrase", ("all",), "train", "text"), "fineweb_edu": ("HuggingFaceFW/fineweb-edu", ("sample-100BT",), "train", "text"), "dclm": ("mlfoundations/dclm-baseline-1.0-parquet", ("default",), "train", "text"), # L3 of openbmb's L0-L4 tiering: Ultra-FineWeb (the L2 selected layer) put # through Q&A generation and multi-style rewriting with MiniCPM4/Qwen3. # Their own decay-phase data for MiniCPM5-1B, which is exactly what we use # it for -- it carries the decay weight that would otherwise go to L2 # general web text. # Math. The eval suite includes ArithMark-3 (arithmetic reasoning) and the # rest of this corpus has no dedicated mathematical content -- an omission, # not a choice. finemath-4plus is the high-quality tier: ~39GB of text # filtered from CommonCrawl by a classifier trained on Llama-3.1-70B # annotations, roughly 8B tokens under our BPE. # 4plus first (the high-quality tier, ~8B tokens under our BPE), then # 3plus as the fallback -- the curriculum draws 10.7B, more than 4plus holds. "finemath": ("HuggingFaceTB/finemath", ("finemath-4plus", "finemath-3plus"), "train", "text"), # 2026-10-10: elementary math facts. Preview-1 sat at chance on ArithMark # division/patterns/multiplication where GPT-X3 (10% UltraData-Math L3 for # its whole run) scores 64-92%. L3 is rewritten math in four styles; one # worker per config so the four stay balanced. "ultra_math_l3": ("openbmb/UltraData-Math", ("UltraData-Math-L3-QA-Synthetic", "UltraData-Math-L3-Textbook-Exercise-Synthetic", "UltraData-Math-L3-Multi-Style-Synthetic", "UltraData-Math-L3-Conversation-Synthetic"), "train", "content"), "ultra_fineweb_l3": ("openbmb/Ultra-FineWeb-L3", ("Ultra-FineWeb-L3-en-QA-Synthetic", "Ultra-FineWeb-L3-en-Multi-Style-Synthetic"), "train", "content"), } NGRAM_N = 13 BAN_CACHE = DATA_DIR / "decontam_ngrams.npy" # --------------------------------------------------------------------------- # _NON_ALNUM = re.compile(r"[^0-9a-z]+") def _norm(s: str) -> list[str]: """Lowercase, split on anything non-alphanumeric. A per-character Python generator here ran once over every document in the corpus -- hundreds of billions of interpreted operations. The regex does the same work in C. """ return _NON_ALNUM.sub(" ", s.lower()).split() def _h(s: str) -> int: """Stable 64-bit hash. Not builtin hash(): PYTHONHASHSEED randomises string hashing per process, so a set built in one process is meaningless in another -- and the ban set is built once and shared by five. 64 bits keeps false positives negligible against ~900k banned n-grams; a 32-bit hash would falsely drop ~0.8% of all documents. """ return int.from_bytes(hashlib.blake2b(s.encode("utf-8", "ignore"), digest_size=8).digest(), "big", signed=True) def eval_ngrams(log=print, use_cache: bool = True): """Sorted array of banned 13-gram hashes, over every evaluation item. Built once and cached: five prepare processes run in parallel and each was rebuilding this from scratch. """ import numpy as _np if use_cache and BAN_CACHE.exists(): a = _np.load(BAN_CACHE) log(f" decontamination: {a.size:,} banned n-grams (cached)") return a import eval as E out = set() missing = [] for name, loader in E.TASKS.items(): try: items = loader() except Exception as e: missing.append(f"{name} ({type(e).__name__}: {str(e)[:60]})") continue n0 = len(out) for it in items: for text in [it["ctx"]] + list(it["choices"]): w = _norm(text) for i in range(max(0, len(w) - NGRAM_N + 1)): out.add(_h(" ".join(w[i:i + NGRAM_N]))) log(f" {name}: {len(items)} items, +{len(out)-n0:,} n-grams") if missing: # Never a warning. A benchmark that silently skips decontamination # produces a score that looks like a result and is not one. raise SystemExit( " decontamination could not load: " + "; ".join(missing) + "\n Refusing to write training data that has not been checked " "against every evaluation set. Fix the loader in eval.py TASKS, or " "pass --no-decontam if you accept scoring on contaminated data.") a = _np.sort(_np.fromiter(out, dtype=_np.int64, count=len(out))) BAN_CACHE.parent.mkdir(parents=True, exist_ok=True) _np.save(BAN_CACHE, a) log(f" decontamination: {a.size:,} banned n-grams -> {BAN_CACHE.name}") return a def contaminated(text: str, ban) -> bool: """True if the document shares a 13-gram window with an evaluation item. The document side steps by NGRAM_N rather than by 1. The evaluation side holds every overlapping window, so any shared span of 2*NGRAM_N-1 = 25 words or more still contains one aligned document window and is always caught -- at a thirteenth of the work. Scanning every overlapping window here was ~160 billion hashes over the full corpus, which does not finish. """ w = _norm(text) if len(w) < NGRAM_N or ban.size == 0: return False keys = np.fromiter( (_h(" ".join(w[i:i + NGRAM_N])) for i in range(0, len(w) - NGRAM_N + 1, NGRAM_N)), dtype=np.int64) if keys.size == 0: return False i = np.searchsorted(ban, keys) i[i >= ban.size] = ban.size - 1 return bool((ban[i] == keys).any()) # --------------------------------------------------------------------------- # class ShardWriter: """Appends uint16 tokens and uint64 document offsets, crash-safely. The offsets file is the authority: on resume the token file is truncated back to the last complete document, so a kill mid-write cannot leave a torn final document that the packer would read as real text. """ def __init__(self, d: Path, eos_id: int): self.d = Path(d); self.d.mkdir(parents=True, exist_ok=True) self.bin = self.d / "tokens.bin" self.idx = self.d / "docs.idx" self.eos = eos_id if not self.idx.exists(): np.array([0], dtype=np.uint64).tofile(self.idx) off = np.fromfile(self.idx, dtype=np.uint64) self.n_tokens = int(off[-1]) self.n_docs = off.size - 1 if self.bin.exists() and self.bin.stat().st_size != self.n_tokens * 2: with open(self.bin, "r+b") as f: f.truncate(self.n_tokens * 2) self.fb = open(self.bin, "ab") self.fi = open(self.idx, "ab") def write(self, docs: list[list[int]]): offs = [] buf = [] n = self.n_tokens for ids in docs: if not ids: continue ids = ids + [self.eos] buf.append(np.asarray(ids, dtype=np.uint16)) n += len(ids) offs.append(n) if not offs: return 0 np.concatenate(buf).tofile(self.fb) np.asarray(offs, dtype=np.uint64).tofile(self.fi) added = n - self.n_tokens self.n_tokens = n self.n_docs += len(offs) return added def flush(self): self.fb.flush(); os.fsync(self.fb.fileno()) self.fi.flush(); os.fsync(self.fi.fileno()) def close(self): self.flush(); self.fb.close(); self.fi.close() def progress_path(name: str) -> Path: return DATA_DIR / name / "progress.json" def load_progress(name: str) -> dict: p = progress_path(name) return json.loads(p.read_text()) if p.exists() else {"seen": 0, "kept": 0, "dropped": 0, "tokens": 0} def save_progress(name: str, d: dict): p = progress_path(name) p.parent.mkdir(parents=True, exist_ok=True) p.write_text(json.dumps(d, indent=2)) # --------------------------------------------------------------------------- # def resolve_files(repo: str, conf: str, split: str) -> list[str]: """The concrete data files behind one config/split, in order. Streaming a whole config and skipping consumed records on resume re-fetches and re-decodes everything already done -- the cost of resuming grows with how far you got, which is the wrong shape for a job that runs for a day. Iterating file by file makes resume O(1): finished files are never touched again. """ from datasets import load_dataset_builder b = load_dataset_builder(repo, conf) df = b.config.data_files files = df[split] if split in df else list(df.values())[0] return [str(f) for f in files] def _stream_file(path: str, field: str): """One data file, one column.""" from datasets import load_dataset fmt = "json" if ".json" in path else "parquet" ds = load_dataset(fmt, data_files=[path], split="train", streaming=True) try: ds = ds.select_columns([field]) # decode one column, not twelve except Exception: pass return ds def prepare_source(name: str, target_tokens: int, cfg: Config, decontam=True, batch=1000, log=print, file_start: int = 0, file_end: int | None = None, out: str | None = None, configs: tuple | None = None): """Tokenize one source into data//. file_start/file_end bound which data files this worker owns. A source whose files are large enough is otherwise one serial process and sets the finish time for the whole job: Ultra-FineWeb needs 106B tokens and was projected at 48h alone while the link and CPU both had headroom. Disjoint file ranges let several workers share it, and because resume is already file-indexed the bookkeeping is just a bound on the loop. data.py stitches data// and data/_s*/ back into one logical source, so the shards are never copied or merged. """ if name not in SOURCES: raise SystemExit(f"unknown source '{name}'. Known: {', '.join(SOURCES)}") repo, confs, split, field = SOURCES[name] if configs: confs = tuple(configs) tok = BorisTokenizer.load(DATA_DIR / "tokenizer") ban = eval_ngrams(log) if decontam else np.zeros(0, dtype=np.int64) out = out or name w = ShardWriter(DATA_DIR / out, tok.eos_id) prog = load_progress(out) # tokens.bin is appended immediately; progress.json is only checkpointed at # file boundaries and every 200M tokens. A worker killed mid-file therefore # has more on disk than its progress claims, and resuming at the recorded # file would write those documents twice. Trust the writer, and skip the # partially-consumed file -- losing the rest of one file is free, duplicate # documents in the corpus are not. if prog.get("tokens", 0) != w.n_tokens: log(f" progress lagged the writer ({prog.get('tokens',0)/1e9:.3f}B " f"recorded vs {w.n_tokens/1e9:.3f}B on disk); resuming past the " f"partial file") prog["tokens"] = w.n_tokens if w.n_tokens: prog["file_idx"] = prog.get("file_idx", file_start) + 1 log(f" {name}: {repo} {list(confs)}") log(f" already {prog['tokens']/1e9:.2f}B tokens in {w.n_docs:,} documents") if w.n_tokens >= target_tokens: log(" target already met"); return t0 = time.time() texts, n_since = [], 0 try: while prog.get("cfg_idx", 0) < len(confs) and prog["tokens"] < target_tokens: conf = confs[prog.get("cfg_idx", 0)] files = resolve_files(repo, conf, split) hi = len(files) if file_end is None else min(file_end, len(files)) if prog.get("file_idx", 0) < file_start: prog["file_idx"] = file_start log(f" config {conf}: {len(files)} files, this worker owns " f"[{file_start}, {hi}), resuming at file {prog.get('file_idx', 0)}") while (prog.get("file_idx", 0) < hi and prog["tokens"] < target_tokens): fi = prog.get("file_idx", 0) # A transient network error must not be fatal. Five of eight workers # died overnight on a single httpx "client has been closed" -- one # WiFi blip -- and sat idle for hours with the job half done. Retry # the file; on persistent failure skip it, because every source here # has far more files than the token budget needs. attempt, rec_iter = 0, None while attempt < 5: try: rec_iter = _stream_file(files[fi], field) break except Exception as e: attempt += 1 log(f" file {fi}: open failed ({type(e).__name__}), " f"retry {attempt}/5 in {2**attempt}s") time.sleep(2 ** attempt) if rec_iter is None: log(f" file {fi}: unreadable after 5 attempts, skipping") prog["file_idx"] = fi + 1 save_progress(out, prog) continue try: for rec in rec_iter: t = rec.get(field) or "" if not t: continue if ban.size and contaminated(t, ban): prog["dropped"] += 1 continue texts.append(t) if len(texts) >= batch: added = w.write(tok.encode_batch(texts)) prog["kept"] += len(texts); prog["tokens"] += added texts = []; n_since += added if n_since > 200_000_000: w.flush(); save_progress(out, prog); n_since = 0 r = prog["tokens"] / max(time.time() - t0, 1e-9) log(f" {prog['tokens']/1e9:7.2f}B / {target_tokens/1e9:.0f}B " f"file {fi+1}/{len(files)} {w.n_docs/1e6:6.2f}M docs " f"{r/1e6:5.2f}M tok/s dropped {prog['dropped']:,} " f"eta {(target_tokens-prog['tokens'])/max(r,1)/3600:5.1f}h") if prog["tokens"] >= target_tokens: break except Exception as e: # Mid-file failure: the documents already written are safe and # the reconciliation on the next start skips the partial file. log(f" file {fi}: stream failed mid-way ({type(e).__name__}: " f"{str(e)[:80]}); moving on") # This file is fully consumed; resume will never fetch it again. prog["file_idx"] = fi + 1 if texts: added = w.write(tok.encode_batch(texts)) prog["kept"] += len(texts); prog["tokens"] += added texts = []; n_since += added w.flush(); save_progress(out, prog) if prog["tokens"] < target_tokens: log(f" config {conf} exhausted at {prog['tokens']/1e9:.2f}B") prog["cfg_idx"] = prog.get("cfg_idx", 0) + 1 prog["file_idx"] = file_start save_progress(out, prog) finally: if texts: prog["tokens"] += w.write(tok.encode_batch(texts)) prog["kept"] += len(texts) w.close(); save_progress(out, prog) log(f" done: {prog['tokens']/1e9:.2f}B tokens, {w.n_docs:,} documents, " f"{prog['dropped']:,} dropped as contaminated") def prepare_sample(mb: int, cfg: Config, log=print): """A stratified text sample for training the tokenizer.""" from datasets import load_dataset out = DATA_DIR / "tokenizer_sample.txt" out.parent.mkdir(parents=True, exist_ok=True) per = mb * 1024 * 1024 // len(SOURCES) with open(out, "w") as f: for name, (repo, confs, split, field) in SOURCES.items(): n = 0 # Not skipped on failure. The tokenizer is the one artifact that # cannot be changed later without invalidating every token file, the # n-gram vocab and every checkpoint, so training it on four of five # sources because one quietly failed is not an acceptable outcome. ds = load_dataset(repo, confs[0], split=split, streaming=True) for rec in ds: t = rec.get(field) or "" if not t: continue f.write(t.replace("\x00", "") + "\n") n += len(t) if n >= per: break log(f" {name}: {n/1e6:.0f} MB") log(f" wrote {out} ({out.stat().st_size/1e6:.0f} MB)") def _val_start(name: str, confs: tuple, log=print): """The (config, file index) where this source's untouched files begin. The old implementation streamed a config from record zero and called next() five million times to get past what training consumed. That re-downloads and re-decodes tens of gigabytes to produce three megabytes of tokens, and with every shard's "seen" counter at 0 it wasn't even reliably disjoint. The download already recorded, per shard, which config and file it stopped at, so the disjoint region is known exactly and reaching it costs nothing. """ shards = [name] + sorted( p.parent.name for p in DATA_DIR.glob(f"{name}_s*/progress.json")) progs = [load_progress(s) for s in shards if progress_path(s).exists()] if not progs: return confs[0], 0 max_cfg = max(p.get("cfg_idx", 0) for p in progs) if max_cfg + 1 < len(confs): return confs[max_cfg + 1], 0 # a whole config training never opened hi = max(p.get("file_idx", 0) for p in progs if p.get("cfg_idx", 0) == max_cfg) return confs[max_cfg], hi + 2 # +2: skip the partially-read file def prepare_val(tokens: int, cfg: Config, log=print): """A fixed held-out split, drawn from every source, that the training packer never opens -- data.open_sources only ever loads cfg.data.sources. Resumable per source: data/val/progress.json records which sources are finished, so an interrupted run does not redo them. """ tok = BorisTokenizer.load(DATA_DIR / "tokenizer") w = ShardWriter(DATA_DIR / "val", tok.eos_id) prog = load_progress("val") done = set(prog.get("done", [])) per = tokens // len(SOURCES) for name, (repo, confs, split, field) in SOURCES.items(): if name in done: log(f" {name}: already in val split"); continue try: conf, fi = _val_start(name, confs, log) files = resolve_files(repo, conf, split) except Exception as e: log(f" {name}: skipped ({type(e).__name__}: {e})"); continue if fi >= len(files): log(f" {name}: no untouched files left in {conf}; skipped") continue got, texts = 0, [] t0 = time.time() while got < per and fi < len(files): try: recs = _stream_file(files[fi], field) except Exception as e: log(f" {name} file {fi}: {type(e).__name__}, skipping") fi += 1; continue for rec in recs: t = rec.get(field) or "" if not t: continue texts.append(t) if len(texts) >= 500: got += w.write(tok.encode_batch(texts)); texts = [] if got >= per: break fi += 1 if texts: got += w.write(tok.encode_batch(texts)) log(f" {name}: {got/1e6:.1f}M val tokens from {conf} " f"(files from {fi}) in {time.time()-t0:.0f}s") done.add(name) save_progress("val", {"done": sorted(done), "tokens": w.n_tokens}) w.close() log(f" val split: {w.n_tokens/1e6:.1f}M tokens, {w.n_docs:,} documents") def check(cfg: Config, log=print): """Open every source and pull one record. Run this before downloading. Every field in SOURCES is a guess until something has actually read from the repo -- two of the five were wrong the first time, and one of them only surfaced after the sample step had already run. """ from datasets import load_dataset bad = 0 for name, (repo, confs, split, field) in SOURCES.items(): for conf in confs: try: ds = load_dataset(repo, conf, split=split, streaming=True) rec = next(iter(ds)) t = rec.get(field) if not isinstance(t, str) or not t: raise ValueError( f"field '{field}' is {type(t).__name__}, not text; " f"available: {list(rec.keys())}") log(f" ok {name:<15}{repo} [{conf}] / {split} / {field} " f"first record {len(t):,} chars") except Exception as e: bad += 1 log(f" FAIL {name:<15}{repo} [{conf}] / {split} / {field}") log(f" {type(e).__name__}: {str(e)[:160]}") if bad: raise SystemExit(f"\n {bad} source/config pairs unreachable -- fix SOURCES " f"before downloading 400GB against a broken registry") log("\n all sources reachable") def plan(cfg: Config, log=print): """Token target per source: what the curriculum will actually draw.""" from schedule import mixture_at total = cfg.train.total_tokens steps = 200 draw = {n: 0.0 for n in cfg.data.sources} means = {} for n in cfg.data.sources: dirs = [d for d in sorted(DATA_DIR.glob(n)) + sorted(DATA_DIR.glob(f"{n}_s*")) if (d / "docs.idx").exists()] nt = nd = 0 for d in dirs: o = np.memmap(d / "docs.idx", dtype=np.uint64, mode="r") nt += int(o[-1]); nd += o.size - 1 means[n] = nt / max(nd, 1) for i in range(steps): w = mixture_at((i + 0.5) / steps, cfg) # The packer samples sources per document, so turn document draw # probabilities into expected token shares before reporting capacity. z = sum(w[n] * means[n] for n in w) or 1.0 for k, v in w.items(): draw[k] += (v * means[k] / z) * total / steps log(f" {cfg.train.total_tokens/1e9:.0f}B token budget, " f"{cfg.total_steps:,} steps\n") log(f" {'source':<16}{'draws':>10}{'unique needed':>16}{'on disk':>12}") log(" " + "-" * 56) for n in cfg.data.sources: have = 0 p = DATA_DIR / n / "docs.idx" if p.exists(): have = int(np.fromfile(p, dtype=np.uint64)[-1]) log(f" {n:<16}{draw[n]/1e9:>9.1f}B{draw[n]/1e9:>15.1f}B" f"{have/1e9:>11.2f}B") log(" " + "-" * 56) log(f" {'TOTAL':<16}{sum(draw.values())/1e9:>9.1f}B" f"{sum(draw.values())/1e9:>15.1f}B") log(f"\n stored as uint16: {sum(draw.values())*2/1e12:.2f} TB at one epoch.") log(f" Repeating a smaller unique corpus is the intended fallback -- set a") log(f" lower --tokens per source and the packer will resample it.") def status(cfg: Config, log=print): log(f" {'source':<16}{'tokens':>12}{'docs':>12}{'dropped':>10}{'GB':>8}") log(" " + "-" * 58) for n in list(cfg.data.sources) + ["val"]: dirs = [d for d in sorted(DATA_DIR.glob(n)) + sorted(DATA_DIR.glob(f"{n}_s*")) if (d / "docs.idx").exists()] if not dirs: log(f" {n:<16}{'-':>12}"); continue tot = docs = dropped = 0 for d in dirs: off = np.fromfile(d / "docs.idx", dtype=np.uint64) tot += int(off[-1]); docs += off.size - 1 dropped += load_progress(d.name).get("dropped", 0) tag = f"{n} ({len(dirs)}sh)" if len(dirs) > 1 else n log(f" {tag:<16}{tot/1e9:>11.2f}B{docs/1e6:>11.2f}M" f"{dropped:>10,}{tot*2/2**30:>8.1f}") def main(): cfg = Config() ap = argparse.ArgumentParser() sub = ap.add_subparsers(dest="cmd", required=True) sub.add_parser("plan"); sub.add_parser("status"); sub.add_parser("check") s = sub.add_parser("sample"); s.add_argument("--mb", type=int, default=2000) s = sub.add_parser("source") s.add_argument("--name", required=True) s.add_argument("--tokens", type=float, required=True) s.add_argument("--no-decontam", action="store_true") s.add_argument("--configs", default=None, help="comma-separated config override, so two workers can " "split a multi-config source between them") s.add_argument("--file-start", type=int, default=0) s.add_argument("--file-end", type=int, default=None) s.add_argument("--out", default=None, help="output directory name; use _sN for a shard") s.add_argument("--batch", type=int, default=250, help="documents per encode_batch call. Each HF Encoding " "carries ids, offsets, attention and word-id arrays, so " "this sets the transient memory peak per worker; 1000 " "across eight workers drove the box 16GB into swap.") s = sub.add_parser("val"); s.add_argument("--tokens", type=float, default=2e7) a = ap.parse_args() if a.cmd == "check": check(cfg) elif a.cmd == "plan": plan(cfg) elif a.cmd == "status": status(cfg) elif a.cmd == "sample": prepare_sample(a.mb, cfg) elif a.cmd == "source": prepare_source(a.name, int(a.tokens), cfg, decontam=not a.no_decontam, file_start=a.file_start, file_end=a.file_end, out=a.out, batch=a.batch, configs=tuple(a.configs.split(",")) if a.configs else None) elif a.cmd == "val": prepare_val(int(a.tokens), cfg) if __name__ == "__main__": main()