File size: 6,217 Bytes
c91c9ec | 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 | #!/usr/bin/env python3
"""RB4 blend builder (runda 3+), zastepuje blend_edu.py z rundy 1.
Poprawki wobec rundy 1 (confoundy z 2026-09-25):
- separator dokumentu brany z tokenizera (<|endoftext|>), nigdy vocab-1;
- zrodlo probkowane z CALEGO pliku na granicach dokumentow (seedowany wybor), nie z glowy;
- manifest z tokenami per zrodlo i per region (dokumenty czatu/QA = zawieraja <|im_start|>).
Tryby:
index --src arcmix.bin --out arcmix_index.npz
granice dokumentow + flaga QA (dokument zawiera <|im_start|>)
sample --src arcmix.bin --index arcmix_index.npz --tokens N [--qa-weight W] --out X.bin
seedowany wybor dokumentow (dokument QA wystepuje W razy w puli) az do N tokenow,
zapis w przetasowanej kolejnosci
mix --a A.bin --b B.bin --a-tokens N --b-tokens M --out X.bin
dwa pliki z tym samym separatorem -> przetasowanie na poziomie dokumentow
Kazdy tryb zapisuje <out>.json (manifest) obok wyjscia.
"""
import argparse
import hashlib
import json
from pathlib import Path
import numpy as np
CHUNK = 250_000_000
def special_ids(tokenizer):
added = {t["content"]: t["id"] for t in json.load(open(tokenizer, encoding="utf-8"))["added_tokens"]}
return added["<|endoftext|>"], added["<|im_start|>"]
def doc_bounds(arr, eos):
"""(starts, lens) dokumentow konczacych sie eos; ogon bez eos pomijany."""
ends = []
for lo in range(0, len(arr), CHUNK):
ends.append(np.flatnonzero(np.asarray(arr[lo:lo + CHUNK]) == eos) + lo)
ends = np.concatenate(ends).astype(np.int64)
starts = np.concatenate([[0], ends[:-1] + 1])
return starts, ends - starts + 1
def write_docs(out_path, parts):
"""parts: iterowalne (array, start, len); zapis + sha256."""
sha, total, buf, buflen = hashlib.sha256(), 0, [], 0
with open(out_path, "wb") as out:
for arr, s, n in parts:
buf.append(np.asarray(arr[s:s + n]))
buflen += n
if buflen >= 50_000_000:
x = np.concatenate(buf)
x.tofile(out)
sha.update(x.tobytes())
total += len(x)
buf, buflen = [], 0
if buf:
x = np.concatenate(buf)
x.tofile(out)
sha.update(x.tobytes())
total += len(x)
return total, sha.hexdigest()
def cmd_index(a):
eos, im_start = special_ids(a.tokenizer)
arr = np.memmap(a.src, dtype=np.uint16, mode="r")
starts, lens = doc_bounds(arr, eos)
qa_pos = np.concatenate([np.flatnonzero(np.asarray(arr[lo:lo + CHUNK]) == im_start) + lo
for lo in range(0, len(arr), CHUNK)])
qa = np.zeros(len(starts), dtype=bool)
qa[np.searchsorted(starts, qa_pos, side="right") - 1] = True
np.savez(a.out, starts=starts, lens=lens, qa=qa)
meta = {"src": a.src, "src_tokens": int(len(arr)), "eos": eos, "docs": int(len(starts)),
"doc_tokens": int(lens.sum()), "qa_docs": int(qa.sum()), "qa_tokens": int(lens[qa].sum())}
Path(a.out + ".json").write_text(json.dumps(meta, indent=1))
print(json.dumps(meta), flush=True)
def cmd_sample(a):
eos, _ = special_ids(a.tokenizer)
arr = np.memmap(a.src, dtype=np.uint16, mode="r")
idx = np.load(a.index)
starts, lens, qa = idx["starts"], idx["lens"], idx["qa"]
pool = np.concatenate([np.arange(len(starts))] + [np.flatnonzero(qa)] * (a.qa_weight - 1))
rng = np.random.default_rng(a.seed)
rng.shuffle(pool)
take = pool[:np.searchsorted(np.cumsum(lens[pool]), a.tokens) + 1]
total, sha = write_docs(a.out, ((arr, int(starts[i]), int(lens[i])) for i in take))
qa_tok = int(lens[take][qa[take]].sum())
meta = {"mode": "sample", "src": a.src, "seed": a.seed, "eos": eos, "qa_weight": a.qa_weight,
"docs": int(len(take)), "unique_docs": int(len(np.unique(take))), "tokens": total,
"qa_tokens": qa_tok, "qa_share": qa_tok / total,
"src_qa_share": float(lens[qa].sum() / lens.sum()), "sha256": sha}
Path(a.out + ".json").write_text(json.dumps(meta, indent=1))
print(json.dumps(meta), flush=True)
def cmd_mix(a):
eos, _ = special_ids(a.tokenizer)
A = np.memmap(a.a, dtype=np.uint16, mode="r")
B = np.memmap(a.b, dtype=np.uint16, mode="r")
sa, la = doc_bounds(A, eos)
sb, lb = doc_bounds(B, eos)
rng = np.random.default_rng(a.seed)
oa, ob = rng.permutation(len(sa)), rng.permutation(len(sb))
oa = oa[:np.searchsorted(np.cumsum(la[oa]), a.a_tokens) + 1]
ob = ob[:np.searchsorted(np.cumsum(lb[ob]), a.b_tokens) + 1]
tagged = np.concatenate([np.stack([np.zeros_like(oa), oa], 1), np.stack([np.ones_like(ob), ob], 1)])
rng.shuffle(tagged)
src = ((A, int(sa[i]), int(la[i])) if t == 0 else (B, int(sb[i]), int(lb[i])) for t, i in tagged)
total, sha = write_docs(a.out, src)
at, bt = int(la[oa].sum()), int(lb[ob].sum())
meta = {"mode": "mix", "a": a.a, "b": a.b, "seed": a.seed, "eos": eos,
"a_docs": int(len(oa)), "a_tokens": at, "b_docs": int(len(ob)), "b_tokens": bt,
"tokens": total, "a_share": at / total, "sha256": sha}
Path(a.out + ".json").write_text(json.dumps(meta, indent=1))
print(json.dumps(meta), flush=True)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--tokenizer", required=True)
ap.add_argument("--seed", type=int, default=1337)
sub = ap.add_subparsers(dest="cmd", required=True)
p = sub.add_parser("index"); p.add_argument("--src", required=True); p.add_argument("--out", required=True)
p = sub.add_parser("sample"); p.add_argument("--src", required=True); p.add_argument("--index", required=True)
p.add_argument("--tokens", type=int, required=True); p.add_argument("--qa-weight", type=int, default=1)
p.add_argument("--out", required=True)
p = sub.add_parser("mix"); p.add_argument("--a", required=True); p.add_argument("--b", required=True)
p.add_argument("--a-tokens", type=int, required=True); p.add_argument("--b-tokens", type=int, required=True)
p.add_argument("--out", required=True)
a = ap.parse_args()
{"index": cmd_index, "sample": cmd_sample, "mix": cmd_mix}[a.cmd](a)
if __name__ == "__main__":
main()
|