File size: 2,610 Bytes
e8eb104 | 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 | #!/usr/bin/env python3
"""Build the fixed token sets every arm is scored on (PROTOCOL.md "Fixed inputs").
uv run --with transformers,pandas python build_eval_set.py --tok <BF16 repo dir> \
--wikitext-test wikitext2-test.txt --c4-eval c4-validation.00001-of-00008.json.gz \
--codeparrot file-000000000054.json.gz --out eval-set.json
Output: {"ppl-wiki": [[ids]*40], "ppl-code": [[ids]*40], "kld": [[ids]*512], "meta": {...}}, all
2048 tokens per window. The same ids go to every arm, so tokenizer drift can't enter the comparison.
Code windows come from the LAST 10 % of codeparrot docs (file order); quant_body.py calibrates on
the first 90 %. C4 eval uses validation shard 1; calibration used shard 0.
"""
import argparse, gzip, hashlib, json, random
from transformers import AutoTokenizer
ap = argparse.ArgumentParser()
ap.add_argument("--tok", required=True); ap.add_argument("--wikitext-test", required=True)
ap.add_argument("--c4-eval", required=True); ap.add_argument("--codeparrot", required=True)
ap.add_argument("--out", required=True)
ap.add_argument("--seqlen", type=int, default=2048)
a = ap.parse_args()
random.seed(608)
tok = AutoTokenizer.from_pretrained(a.tok)
L = a.seqlen
def windows(texts, n):
out, buf = [], []
for t in texts:
buf += tok(t, add_special_tokens=False)["input_ids"] + [tok.eos_token_id]
while len(buf) >= L and len(out) < n:
out.append(buf[:L]); buf = buf[L:]
if len(out) >= n:
return out
raise SystemExit(f"only {len(out)} windows, wanted {n}")
wiki_ids = tok(open(a.wikitext_test, encoding="utf-8").read(), add_special_tokens=False)["input_ids"]
ppl_wiki = [wiki_ids[i * L:(i + 1) * L] for i in range(40)] # same chunks as ppl_wikitext2.py
code_docs = [json.loads(l)["content"] for l in gzip.open(a.codeparrot, "rt")]
code_tail = code_docs[int(len(code_docs) * 0.9):]
random.shuffle(code_tail)
code_win = windows(code_tail, 256 + 40)
ppl_code, kld_code = code_win[:40], code_win[40:]
c4 = [json.loads(l)["text"] for l in gzip.open(a.c4_eval, "rt")]
random.shuffle(c4)
kld = windows(c4, 256) + kld_code
out = {"ppl-wiki": ppl_wiki, "ppl-code": ppl_code, "kld": kld}
blob = json.dumps(out, separators=(",", ":")).encode()
out["meta"] = {"seqlen": L, "seed": 608, "tokenizer": a.tok, "counts": {k: len(v) for k, v in out.items()},
"sha256_ids": hashlib.sha256(blob).hexdigest(), "code_docs_total": len(code_docs),
"code_tail_from_index": int(len(code_docs) * 0.9)}
json.dump(out, open(a.out, "w"), separators=(",", ":"))
print(json.dumps(out["meta"], indent=1))
|