Datasets:
File size: 8,013 Bytes
c9faf30 | 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 | """Move the flat DeepSeek-era corpus into the pool, losslessly.
python tools/pool_import.py
`calib_train.txt` was written by `pipeline/build.py` as documents joined with a
blank line, in ascending `id` order, and `calib_train.manifest.jsonl` records
each document's character count in that same order. That makes the split exact
rather than heuristic: this script asserts the reconstruction re-joins to the
original file byte for byte and refuses to write anything if it does not.
Nothing at the repository root is modified. `calib_heldout.txt` is deliberately
*not* imported: it is a measurement split, and putting it in the pool would let
a later build calibrate on it.
"""
from __future__ import annotations
import argparse
import json
import os
import subprocess
import sys
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import poollib as P
# legacy domain -> (pool category, subdirectory, shard name, synthetic)
DOMAIN_MAP = {
"code": ("code", "", "legacy-dsv4", False),
"graphics": ("graphics", "", "legacy-dsv4", False),
"structured": ("structured", "", "legacy-dsv4", False),
"longctx": ("longctx", "", "legacy-dsv4", False),
"general": ("multilingual", "", "legacy-dsv4-wikipedia", False),
"agentic": ("agentic", "synthetic", "legacy-dsv4-rendered", True),
"reasoning": ("reasoning", "synthetic", "legacy-dsv4-rendered", True),
"vocab_sweep": ("vocab_sweep", "synthetic", "legacy-dsv4-rendered", True),
}
# Domains whose text is DeepSeek chat markup. Kept for provenance, never
# usable as calibration for another model: the markup would tokenise as literal
# punctuation and calibrate nothing the target model ever emits.
DSV4_RENDERED = {"agentic", "reasoning"}
# vocab_sweep is a function of the tokenizer, not a corpus. The DeepSeek sweep
# is kept but a build regenerates its own.
TOKENIZER_BOUND = {"vocab_sweep"}
def split_flat(txt_path: str, manifest_path: str) -> list[tuple[dict, str]]:
recs = [json.loads(l) for l in open(manifest_path, encoding="utf-8") if l.strip()]
if recs != sorted(recs, key=lambda r: r["id"]):
raise SystemExit(f"{manifest_path}: not in ascending id order; "
"the offset reconstruction assumes pipeline/build.py ordering")
text = open(txt_path, encoding="utf-8").read()
out, pos = [], 0
for i, r in enumerate(recs):
doc = text[pos:pos + r["chars"]]
pos += r["chars"]
if i < len(recs) - 1:
if text[pos:pos + 2] != "\n\n":
raise SystemExit(f"{txt_path}: expected blank-line separator after "
f"document {i} ({r['id']}), found {text[pos:pos+4]!r}")
pos += 2
# the writer appended a single trailing newline
out.append((r, doc))
if text[pos:] != "\n":
raise SystemExit(f"{txt_path}: unexpected trailing bytes {text[pos:][:40]!r}")
rejoined = "\n\n".join(d for _, d in out) + "\n"
if rejoined != text:
raise SystemExit(f"{txt_path}: reconstruction is not byte-identical")
return out
def git_sha(repo_root: str) -> str:
try:
return subprocess.run(["git", "-C", repo_root, "rev-parse", "HEAD"],
capture_output=True, text=True, check=True).stdout.strip()
except Exception:
return "unknown"
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--repo", default=P.REPO_ROOT)
args = ap.parse_args()
root = args.repo
sha = git_sha(root)
train = split_flat(os.path.join(root, "calib_train.txt"),
os.path.join(root, "calib_train.manifest.jsonl"))
print(f"calib_train.txt -> {len(train)} documents, reconstruction byte-identical")
buckets: dict[str, list[P.Record]] = {}
unmapped: list[dict] = []
for r, text in train:
dom = r["domain"]
if dom not in DOMAIN_MAP:
unmapped.append(r)
continue
cat, sub, shard, synth = DOMAIN_MAP[dom]
render = "dsv4" if dom in DSV4_RENDERED else "text"
prov = {
"imported_from": "calib_train.txt",
"import_commit": sha,
"legacy_domain": dom,
"legacy_content_domain": r.get("content_domain", dom),
"legacy_id": r["id"],
"legacy_tokens_deepseek_v4": r.get("tokens", 0),
"legacy_split": r.get("split", "calib_train"),
}
if dom in DSV4_RENDERED:
prov["excluded_from_builds"] = (
"text is rendered in DeepSeek-V4 chat markup (<|DSML|...>, "
"<|User|>, <think>); those byte sequences are not special tokens "
"for any other tokenizer. The equivalent conversations are "
"regenerated structurally by tools/gen_agentic.py and "
"tools/gen_reasoning.py.")
if dom in TOKENIZER_BOUND:
prov["excluded_from_builds"] = (
"a vocabulary sweep is a function of the tokenizer it was built "
"for (DeepSeek-V4, 129,280 rows). tools/vocab_sweep.py "
"regenerates it for the target model.")
rec = P.Record(
category=cat, domain=dom, source=r["source"], license=r["license"],
license_url="", path=r["path"], lang=r["lang"], origin=r["origin"],
text=text, synthetic=synth, render=render, provenance=prov,
)
key = os.path.join(cat, sub, shard + ".jsonl") if sub else os.path.join(cat, shard + ".jsonl")
buckets.setdefault(key, []).append(rec)
if unmapped:
# Never dropped: anything without a mapping lands in pool/unsorted with
# the reason attached, so it can be triaged rather than lost.
recs = []
for r, text in train:
if r in unmapped:
recs.append(P.Record(
category="unsorted", domain=r["domain"], source=r["source"],
license=r["license"], path=r["path"], lang=r["lang"],
origin=r["origin"], text=text,
provenance={"imported_from": "calib_train.txt",
"import_commit": sha,
"reason": f"legacy domain {r['domain']!r} has no pool category"},
))
buckets["unsorted/legacy-dsv4-unmapped.jsonl"] = recs
total = 0
for key in sorted(buckets):
n = P.write_jsonl(os.path.join(P.POOL_ROOT, key), buckets[key])
chars = sum(len(r.text) for r in buckets[key])
total += n
print(f" pool/{key:52} {n:5} docs {chars:9,} chars")
print(f" {total} documents written, {len(unmapped)} unmapped")
# eval_neutral is copied into eval/neutral with provenance attached; the
# file at the repository root is left exactly as it is because the existing
# measurements are keyed to it.
ev = split_flat(os.path.join(root, "eval_neutral.txt"),
os.path.join(root, "eval_neutral.manifest.jsonl"))
erecs = [P.Record(
category="code", domain=r["domain"], source=r["source"],
license=r["license"], license_url="", path=r["path"], lang=r["lang"],
origin=r["origin"], text=text,
provenance={"imported_from": "eval_neutral.txt", "import_commit": sha,
"legacy_id": r["id"],
"note": "the flat file was renamed to eval/code/eval_code.txt and kept "
"byte-identical; measurement showed it is a code corpus, not "
"neutral prose. This is a categorised copy with provenance"},
) for r, text in ev]
n = P.write_jsonl(os.path.join(P.EVAL_ROOT, "code", "eval-code-legacy.jsonl"), erecs)
print(f"eval_neutral.txt -> eval/code/eval-code-legacy.jsonl {n} docs")
return 0
if __name__ == "__main__":
sys.exit(main())
|