Blame for math/anneal/code/prep_text.py

"""SHADOW Assistant text data: download, route, tokenise (Gemma 4 vocabulary of EmbeddingGemma 2), pack 3 bytes per token,

upload shards as they finish (local disk only holds the shards in flight). Each shard also saves its unigram counts

(.counts.npz), so the frozen table's log-unigram prior is a sum over these files, with no second pass over the data.



    python prep_text.py --source fineweb_edu --workers 30

    python prep_text.py --source legal --test           # tiny local check: 2 files, 2M tokens per bucket, no upload



Layout, one folder per source:  <area>/<source>/<source>_w<worker>_<shard>.bin (+ .json, .counts.npz)

  fineweb_edu   HuggingFaceFW/fineweb-edu: .au pages -> knowledge/fineweb_edu_au (all), others -> knowledge/fineweb_edu (75B)

  dclm          mlfoundations/dclm-baseline-1.0: .au pages -> knowledge/dclm_au (6B, skipping URLs already in FineWeb-Edu .au),

                others -> knowledge/dclm (35B)

  legal, spaceai, joeyllm   whole dataset -> knowledge/<source>

  opencoder_fineweb, opencoder_annealing   whole dataset -> code/<source>

  nemotron_math  nvidia/Nemotron-CC-Math-v1 4plus -> math/nemotron_cc_math_4plus, private repo (training only, rebuilt not posted)

  wikipedia_au  Australian articles of English Wikipedia -> knowledge/wikipedia_au, private repo (CC BY-SA, rebuilt not posted)

  anneal_*      maths for the anneal (SK 2026-10-23) -> math/anneal/<source>: Dolmino-mix-1124 maths subsets (MathCoder2,

                TuluMath, DolminoSynthMath, GSM8K train, CodeSearchNet-owmfilter), MegaMath Q&A, and MegaMath Q&A

Shard file: raw tokens, 3 bytes each little-endian, every document = <bos> text <eos>. The .json lists tokens, documents

and source files.

"""
import argparse, hashlib, json, multiprocessing as mp, os, pathlib, re, time
from urllib.parse import urlparse
import numpy as np

BOS, EOS, VOCAB = 2, 1, 262144
SRC = {
    "fineweb_edu": {"repo": "HuggingFaceFW/fineweb-edu", "glob": r"^data/.*\.parquet$", "route": "au_or_global",
                    "au": "knowledge/fineweb_edu_au", "glb": "knowledge/fineweb_edu", "glb_quota": 75e9, "au_quota": 1e12, "save_au_urls": True},
    "dclm": {"repo": "mlfoundations/dclm-baseline-1.0", "glob": r"\.jsonl\.zst$", "route": "au_or_global",
             "au": "knowledge/dclm_au", "glb": "knowledge/dclm", "glb_quota": 35e9, "au_quota": 6e9, "skip_au_urls": True},
    "dclm_global": {"repo": "mlfoundations/dclm-baseline-1.0", "glob": r"\.jsonl\.zst$", "route": "au_or_global",          # global half only
                    "au": "knowledge/dclm_au", "glb": "knowledge/dclm", "glb_quota": 35e9, "au_quota": 0},
    "dclm_au": {"repo": "mlfoundations/dclm-baseline-1.0", "glob": r"\.jsonl\.zst$", "route": "au_or_global",              # .au half only
                "au": "knowledge/dclm_au", "glb": "knowledge/dclm", "glb_quota": 0, "au_quota": 6e9, "skip_au_urls": True},
    "legal": {"repo": "isaacus/open-australian-legal-corpus", "glob": r"\.(parquet|jsonl|jsonl\.gz)$", "route": "all", "all": "knowledge/legal", "quota": 1e12},
    "spaceai": {"repo": "MrSpaceCommander/spaceai-australian-corpus", "glob": r"^data/(train|validation)/(?!oalc).*[.]jsonl[.]gz$",   # its oalc files repeat the legal corpus
                "route": "all", "all": "knowledge/spaceai", "quota": 1e12},
    "joeyllm": {"repo": "JoeyLLM/australian-dataset-1b", "glob": r"\.parquet$", "route": "all", "all": "knowledge/joeyllm", "quota": 1e12, "skip_au_urls": True},
    "opencoder_fineweb": {"repo": "OpenCoder-LLM/opc-fineweb-code-corpus", "glob": r"^data/.*\.parquet$", "route": "all", "all": "code/opencoder_fineweb", "quota": 1e12},
    "opencoder_annealing": {"repo": "OpenCoder-LLM/opc-annealing-corpus", "glob": r"\.jsonl$", "route": "all", "all": "code/opencoder_annealing", "quota": 1e12},
    "gutenberg": {"repo": "juniot/qualc-gutenberg-en", "glob": r"^data/.*\.parquet$", "route": "all", "all": "creative/gutenberg", "quota": 1e12},
    "poetry_lines": {"repo": "biglam/gutenberg-poetry-corpus", "glob": r"^data/.*\.parquet$", "route": "all", "all": "creative/gutenberg_poetry",
                     "quota": 1e12, "text": "poetry_lines"},
    "poems": {"repo": "DanFosing/public-domain-poetry", "glob": r"^poems\.json$", "route": "all", "all": "creative/public_domain_poetry", "quota": 1e12, "text": "poem"},
    "cosmopedia_stories": {"repo": "HuggingFaceTB/cosmopedia", "glob": r"^data/stories/.*\.parquet$", "route": "all", "all": "creative/cosmopedia_stories", "quota": 1e12},
    "nemotron_math": {"repo": "nvidia/Nemotron-CC-Math-v1", "glob": r"^4plus/.*\.parquet$", "route": "all", "all": "math/nemotron_cc_math_4plus",
                      "quota": 1e12, "private": True},                     # NVIDIA's terms: training only, rebuilt not posted
    "wikipedia_au": {"repo": "wikimedia/wikipedia", "glob": r"^20231101\.en/.*\.parquet$", "route": "all", "all": "knowledge/wikipedia_au",
                     "quota": 1e12, "filter": "australia", "private": True},
    "anneal_mathcoder2": {"repo": "allenai/dolmino-mix-1124", "glob": r"^data/math/mathcoder2-synthmath/.*\.jsonl(\.gz|\.zst)?$", "route": "all", "all": "math/anneal/mathcoder2", "quota": 1e12},
    "anneal_tulu_math": {"repo": "allenai/dolmino-mix-1124", "glob": r"^data/math/tulu_math/.*\.jsonl(\.gz|\.zst)?$", "route": "all", "all": "math/anneal/tulu_math", "quota": 1e12},
    "anneal_dolmino_synth": {"repo": "allenai/dolmino-mix-1124", "glob": r"^data/math/dolmino_math_synth/.*\.jsonl(\.gz|\.zst)?$", "route": "all", "all": "math/anneal/dolmino_math_synth", "quota": 1e12},
    "anneal_gsm8k": {"repo": "allenai/dolmino-mix-1124", "glob": r"^data/math/gsm8k/main/train/.*\.jsonl(\.gz|\.zst)?$", "route": "all", "all": "math/anneal/gsm8k", "quota": 1e12},
    "anneal_codesearchnet": {"repo": "allenai/dolmino-mix-1124", "glob": r"^data/math/codesearchnet-owmfilter/.*\.jsonl(\.gz|\.zst)?$", "route": "all", "all": "math/anneal/codesearchnet_owmfilter", "quota": 1e12},
    "anneal_megamath_qa": {"repo": "LLM360/MegaMath", "glob": r"^megamath-qa/.*\.parquet$", "route": "all", "all": "math/anneal/megamath_qa", "quota": 1e12},
}
AU_WORDS = re.compile(r"\b(Australia|Australian|New South Wales|Victoria, Australia|Queensland|Tasmania|Western Australia|South Australia|Northern Territory|Canberra|Sydney|Melbourne|Brisbane|Perth|Adelaide|Hobart|Darwin)\b")

def is_au(url):
    try: return urlparse(url).netloc.lower().split(":")[0].endswith(".au")
    except Exception: return False

def rows_jsonl_zst(repo, f):
    """read a .jsonl.zst file line by line, keeping only text and url (DCLM files disagree on the types of other fields)"""
    import io, zstandard
    from huggingface_hub import HfFileSystem
    with HfFileSystem().open(f"datasets/{repo}/{f}", "rb", block_size=16 << 20) as fh:
        for line in io.TextIOWrapper(zstandard.ZstdDecompressor().stream_reader(fh), encoding="utf-8", errors="replace"):
            try: o = json.loads(line)
            except Exception: continue
            yield {"text": o.get("text") or "", "url": o.get("url") or ""}

def poems_from_lines(rows, max_chars=20000):
    """the Gutenberg Poetry Corpus has one row per line; join consecutive lines of the same book into documents"""
    cur, lines, n = None, [], 0
    for r in rows:
        if (r["gutenberg_id"] != cur or n > max_chars) and lines: yield {"text": "\n".join(lines)}; lines, n = [], 0
        cur = r["gutenberg_id"]; lines.append(r["line"]); n += len(r["line"]) + 1
    if lines: yield {"text": "\n".join(lines)}

def uhash(url): return hashlib.blake2b(url.encode("utf-8", "ignore"), digest_size=8).digest()

class Shards:
    """collects tokens for one bucket and writes 3-byte shards; finished shard names go to the upload queue"""
    def __init__(self, out, bucket, wid, shard_tokens, upq):
        self.out, self.bucket, self.wid, self.shard_tokens, self.upq = out, bucket, wid, shard_tokens, upq
        self.buf, self.n, self.docs, self.k, self.files = [], 0, 0, 0, set()
    def add(self, ids_list, src_file):
        for ids in ids_list:
            self.buf.append(np.asarray([BOS] + ids + [EOS], np.uint32)); self.n += len(ids) + 2; self.docs += 1
        self.files.add(src_file)
        if self.n >= self.shard_tokens: self.flush()
    def flush(self):
        if not self.n: return
        a = np.concatenate(self.buf); b = np.empty((len(a), 3), np.uint8); b[:, 0] = a & 255; b[:, 1] = (a >> 8) & 255; b[:, 2] = (a >> 16) & 255
        name = f"{self.bucket}/{self.bucket.split('/')[-1]}_w{self.wid:02d}_{self.k:05d}"
        p = pathlib.Path(self.out) / (name + ".bin"); p.parent.mkdir(parents=True, exist_ok=True); p.write_bytes(b.tobytes())
        np.savez_compressed(pathlib.Path(self.out) / (name + ".counts.npz"), counts=np.bincount(a, minlength=VOCAB).astype(np.uint32))
        meta = {"tokens": int(len(a)), "documents": self.docs, "bytes_per_token": 3, "bos": BOS, "eos": EOS, "source_files": sorted(self.files)}
        pathlib.Path(self.out, name + ".json").write_text(json.dumps(meta))
        self.upq.put(name); self.buf, self.n, self.docs, self.files = [], 0, 0, set(); self.k += 1

def worker(wid, files, args, cfg, counters, upq, ret):
    from datasets import load_dataset
    from transformers import AutoTokenizer
    tk = AutoTokenizer.from_pretrained("google/embeddinggemma-2")
    skip = set()
    if cfg.get("skip_au_urls") and os.path.exists(args.au_urls):
        raw = pathlib.Path(args.au_urls).read_bytes(); skip = {raw[i:i + 8] for i in range(0, len(raw), 8)}
    if cfg.get("skip_urls_file"):                                                  # pages already in another source (by URL)
        raw = pathlib.Path(cfg["skip_urls_file"]).read_bytes(); skip = {raw[i:i + 8] for i in range(0, len(raw), 8)}
    au_seen = [] if cfg.get("save_au_urls") else None
    sh = {}
    def S(b):
        if b not in sh: sh[b] = Shards(args.out, b, wid, args.shard_tokens, upq); sh[b].k = args.shard_start
        return sh[b]
    def room(key, quota):
        with counters[key].get_lock(): return counters[key].value < quota
    quotas = [("glb", cfg["glb_quota"]), ("au", cfg["au_quota"])] if cfg["route"] == "au_or_global" else [("all", cfg["quota"])]
    def put(batch, f):
        for (bb, kk, qq), texts in batch.items():
            if not texts or not room(kk, qq): continue
            ids = tk(texts, add_special_tokens=False)["input_ids"]; S(bb).add(ids, f)
            with counters[kk].get_lock(): counters[kk].value += sum(len(i) + 2 for i in ids)
    for fi, f in enumerate(files):
        if all(not room(k, q) for k, q in quotas): break
        ds = rows_jsonl_zst(cfg["repo"], f) if f.endswith(".jsonl.zst") else load_dataset(cfg["repo"], data_files=[f], split="train", streaming=True)
        if cfg.get("text") == "poetry_lines": ds = poems_from_lines(ds)
        batch = {}
        for row in ds:
            t = row.get("text") or ""
            if cfg.get("text") == "poem": t = f"{row.get('Title') or ''}\nby {row.get('Author') or 'Anonymous'}\n\n{t}".strip()
            if len(t) < 50: continue
            if cfg.get("filter") == "australia" and not AU_WORDS.search(t[:3000]): continue
            if cfg["route"] == "au_or_global":
                url = row.get("url") or ""
                if is_au(url):
                    h = uhash(url)
                    if h in skip: continue
                    if au_seen is not None: au_seen.append(h)
                    key = (cfg["au"], "au", cfg["au_quota"])
                else: key = (cfg["glb"], "glb", cfg["glb_quota"])
            else:
                if (cfg.get("skip_au_urls") or cfg.get("skip_urls_file")) and uhash(row.get("url") or "") in skip: continue
                key = (cfg["all"], "all", cfg["quota"])
            batch.setdefault(key, []).append(t)
            if sum(len(v) for v in batch.values()) >= 512:
                put(batch, f); batch = {}
                if all(not room(k, q) for k, q in quotas): break          # every bucket of this source is full: stop reading
        put(batch, f)
        print(f"w{wid:02d} file {fi + 1}/{len(files)} done {f} | totals " + " ".join(f"{k}={v.value / 1e9:.3f}B" for k, v in counters.items()), flush=True)
    for s in sh.values(): s.flush()
    if au_seen: ret.put(b"".join(au_seen))

def upload(names, out, repo):
    """one commit per source folder for a batch of finished shards, then delete the local files"""
    from huggingface_hub import HfApi, CommitOperationAdd
    api = HfApi(); groups = {}
    for n in names: groups.setdefault(n.rsplit("/", 1)[0], []).append(n)
    for folder, ns in groups.items():
        ops = [CommitOperationAdd(path_in_repo=n + ext, path_or_fileobj=str(pathlib.Path(out) / (n + ext))) for n in ns for ext in (".bin", ".json", ".counts.npz")]
        for attempt in range(8):
            try: api.create_commit(repo, repo_type="dataset", operations=ops, commit_message=f"Add {folder} ({len(ns)} shards)"); break
            except Exception as e: print(f"upload retry {attempt + 1} {folder}: {str(e)[:150]}", flush=True); time.sleep(30 * (attempt + 1))
        else: print(f"UPLOAD FAILED {folder}: kept on disk", flush=True); continue
        for n in ns:
            for ext in (".bin", ".json", ".counts.npz"): (pathlib.Path(out) / (n + ext)).unlink(missing_ok=True)

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--source", required=True, choices=list(SRC)); ap.add_argument("--workers", type=int, default=max(1, os.cpu_count() - 4))
    ap.add_argument("--out", default="out"); ap.add_argument("--repo", default="QLNI/shadow-assistant-data")
    ap.add_argument("--private-repo", default="QLNI/shadow-assistant-private"); ap.add_argument("--shard-tokens", type=int, default=250_000_000)
    ap.add_argument("--batch", type=int, default=8, help="shards per commit"); ap.add_argument("--au-urls", default="fineweb_edu_au_urls.bin")
    ap.add_argument("--test", action="store_true"); ap.add_argument("--skip-files", default="", help="file listing source files already done")
    ap.add_argument("--glb-quota", type=float, default=0); ap.add_argument("--shard-start", type=int, default=0, help="first shard number (resume)")
    args = ap.parse_args(); cfg = dict(SRC[args.source]); repo = args.private_repo if cfg.get("private") else args.repo
    if args.glb_quota: cfg["glb_quota"] = args.glb_quota
    from huggingface_hub import HfApi
    files = sorted(f for f in HfApi().list_repo_files(cfg["repo"], repo_type="dataset") if re.search(cfg["glob"], f))
    if args.skip_files:
        done = set(open(args.skip_files).read().split()); files = [f for f in files if f not in done]
    if args.test:
        files = files[:2]; args.workers = min(2, len(files)); args.shard_tokens = 1_000_000; repo = ""
        for k in ("glb_quota", "au_quota", "quota"):
            if k in cfg: cfg[k] = 2e6
    print(f"{args.source}: {len(files)} files, {args.workers} workers, repo {repo or '(none, local only)'}", flush=True)
    counters = {k: mp.Value("d", 0.0) for k in (["glb", "au"] if cfg["route"] == "au_or_global" else ["all"])}
    upq, ret = mp.Queue(), mp.Queue(); chunks = [files[i::args.workers] for i in range(args.workers)]
    t0 = time.time(); ps = [mp.Process(target=worker, args=(i, chunks[i], args, cfg, counters, upq, ret)) for i in range(args.workers)]
    for p in ps: p.start()
    pending, urls, last = [], [], time.time()
    def drain():
        while True:
            try: pending.append(upq.get_nowait())
            except Exception: break
        while True:
            try: urls.append(ret.get_nowait())
            except Exception: break
    while any(p.is_alive() for p in ps):                # the main process does all uploads, in batches
        time.sleep(5); drain()
        if repo and (len(pending) >= args.batch or (pending and time.time() - last > 900)):
            upload(pending, args.out, repo); pending, last = [], time.time()
    for p in ps: p.join()
    drain()
    if repo and pending: upload(pending, args.out, repo)
    if cfg.get("save_au_urls") and urls: pathlib.Path(args.au_urls).write_bytes(b"".join(urls))
    summary = {"source": args.source, "files": len(files), "hours": round((time.time() - t0) / 3600, 3), **{k: v.value for k, v in counters.items()}}
    print("SUMMARY", json.dumps(summary), flush=True)
    pathlib.Path(args.out).mkdir(exist_ok=True); pathlib.Path(args.out, f"summary_{args.source}.json").write_text(json.dumps(summary, indent=1))

if __name__ == "__main__":
    main()