File size: 8,312 Bytes
5952424 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | """Build per-fold memmap shards from a 10-fold_split fold's train.jsonl.zst.
Same output format/planes as build_shards.py / build_bronze.py so the training loader
works unchanged when STOICHEIA_DATA points at the fold root:
<out>/v1_punct/ tier in {pristine, repaired} (pristine rows first, then
repaired — mirrors the flagship's canonical order so the
idx%HOLDOUT_MOD holdout and eval/intrinsic sampling behave
identically)
<out>/bronze_punct/ tier == bronze
Records with tier == inscriptions (or anything else) are skipped and counted — the
flagship recipe never trains on inscriptions. Record ids (including the #segN suffixes
of excised-and-restitched train segments) are kept verbatim.
Usage:
python data/build_fold_shards.py --jsonl .../fold_0/train.jsonl.zst \
--out $STOICHEIA_DATA/folds/fold_0/shards --workers 16
"""
from __future__ import annotations
import argparse, json, shutil, subprocess, sys
from collections import Counter
from concurrent.futures import ProcessPoolExecutor
from dataclasses import asdict
from pathlib import Path
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from data.normalize import Stats, normalize_record
PLANES = ("chars", "boundary", "dia", "cap", "punct")
TIER2TARGET = {"pristine": "pri", "repaired": "rep", "bronze": "brz"}
CLEAN = {"pristine": 1.0, "repaired": 0.75, "bronze": 0.5}
INDEX_SCHEMA = pa.schema([("offset", pa.int64()), ("length", pa.int64()),
("tier", pa.string()), ("clean", pa.float64()),
("source", pa.string()), ("id", pa.string())])
def _dump_stats(stats: Stats, path: Path):
st = asdict(stats)
st["stripped"] = dict(stats.stripped)
st["archaic"] = dict(stats.archaic)
st["other_marks"] = dict(stats.other_marks)
path.write_text(json.dumps(st))
def _load_stats(path: Path) -> Stats:
s = json.loads(path.read_text())
st = Stats(**{k: s[k] for k in ("records_in", "records_kept",
"records_dropped_nongreek", "records_dropped_empty", "letters",
"words", "sentences", "mark_conflicts", "orphan_marks")})
st.stripped = Counter({int(k): v for k, v in s["stripped"].items()})
st.archaic = Counter(s["archaic"])
st.other_marks = Counter(s["other_marks"])
return st
def process_chunk(args):
chunk_file, tmpdir, cid = args
tmpdir = Path(tmpdir)
stats = {t: Stats() for t in ("pri", "rep", "brz")}
bufs = {t: {p: [] for p in PLANES} for t in ("pri", "rep", "brz")}
rows = {t: {"offset": [], "length": [], "tier": [], "clean": [], "source": [], "id": []}
for t in ("pri", "rep", "brz")}
off = {t: 0 for t in ("pri", "rep", "brz")}
skipped = Counter()
with open(chunk_file) as f:
for line in f:
try:
rec = json.loads(line)
except Exception:
skipped["unparseable"] += 1
continue
tier = rec.get("tier", "")
t = TIER2TARGET.get(tier)
if t is None:
skipped[tier or "missing_tier"] += 1
continue
r = normalize_record(rec.get("text", ""), stats[t], with_punct=True)
if r is None:
continue
for p, a in zip(PLANES, r):
bufs[t][p].append(a)
rows[t]["offset"].append(off[t])
rows[t]["length"].append(len(r[0]))
rows[t]["tier"].append(tier)
rows[t]["clean"].append(CLEAN[tier])
rows[t]["source"].append(rec.get("source", ""))
rows[t]["id"].append(rec.get("id", ""))
off[t] += len(r[0])
for t in ("pri", "rep", "brz"):
d = tmpdir / t
d.mkdir(parents=True, exist_ok=True)
for p in PLANES:
if bufs[t][p]:
np.concatenate(bufs[t][p]).tofile(d / f"{p}.bin")
else:
np.array([], np.uint8).tofile(d / f"{p}.bin")
pq.write_table(pa.table(rows[t], schema=INDEX_SCHEMA), d / "index.parquet")
_dump_stats(stats[t], d / "stats.json")
(tmpdir / "skipped.json").write_text(json.dumps(dict(skipped)))
return cid
def assemble(out_dir: Path, part_dirs: list[Path]):
"""Concatenate per-chunk target dirs (in the given order) into one shard dir."""
out_dir.mkdir(parents=True, exist_ok=True)
total = Stats()
tables, cum = [], 0
for d in part_dirs:
t = pq.read_table(d / "index.parquet")
if t.num_rows:
t = t.set_column(0, "offset", pa.array(t.column("offset").to_numpy() + cum,
type=pa.int64()))
tables.append(t)
cum += int(np.sum(t.column("length").to_numpy()))
total.merge(_load_stats(d / "stats.json"))
for p in PLANES:
with open(out_dir / f"{p}.bin", "wb") as fo:
for d in part_dirs:
fo.write((d / f"{p}.bin").read_bytes())
pq.write_table(pa.concat_tables(tables) if tables else INDEX_SCHEMA.empty_table(),
out_dir / "index.parquet")
st = asdict(total)
st["stripped"] = {str(k): v for k, v in total.stripped.items()}
st["archaic"] = dict(total.archaic)
st["other_marks"] = dict(total.other_marks)
(out_dir / "stats.json").write_text(json.dumps(st, indent=2, ensure_ascii=False))
return cum, total
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--jsonl", required=True, help="fold_k/train.jsonl.zst")
ap.add_argument("--out", required=True, help="fold root shards dir (gets v1_punct/, bronze_punct/)")
ap.add_argument("--workers", type=int, default=16)
ap.add_argument("--chunk", type=int, default=40000)
a = ap.parse_args()
out = Path(a.out)
out.mkdir(parents=True, exist_ok=True)
tmp = out / "tmp"
if tmp.exists():
shutil.rmtree(tmp)
(tmp / "chunks").mkdir(parents=True)
# stream-decompress and split into line-chunk files (keeps memory flat)
proc = subprocess.Popen(["zstdcat", a.jsonl], stdout=subprocess.PIPE, text=True)
chunks, buf, cid = [], [], 0
for line in proc.stdout:
buf.append(line)
if len(buf) >= a.chunk:
cf = tmp / "chunks" / f"c{cid:04d}.jsonl"
cf.write_text("".join(buf))
chunks.append((str(cf), str(tmp / f"c{cid:04d}"), cid))
buf, cid = [], cid + 1
if buf:
cf = tmp / "chunks" / f"c{cid:04d}.jsonl"
cf.write_text("".join(buf))
chunks.append((str(cf), str(tmp / f"c{cid:04d}"), cid))
if proc.wait() != 0:
raise RuntimeError(f"zstdcat failed on {a.jsonl}")
print(f"{len(chunks)} chunks", flush=True)
with ProcessPoolExecutor(max_workers=a.workers) as ex:
list(ex.map(process_chunk, chunks))
order = [Path(c[1]) for c in chunks]
# canonical order: all pristine parts, then all repaired parts (mirrors build_shards.py)
lit_letters, lit_stats = assemble(out / "v1_punct",
[d / "pri" for d in order] + [d / "rep" for d in order])
brz_letters, brz_stats = assemble(out / "bronze_punct", [d / "brz" for d in order])
skipped = Counter()
for d in order:
skipped.update(json.loads((d / "skipped.json").read_text()))
prov = {
"source_jsonl": str(Path(a.jsonl).resolve()),
"source_mtime": Path(a.jsonl).stat().st_mtime,
"v1_punct": {"letters": lit_letters, "records_kept": lit_stats.records_kept,
"records_in": lit_stats.records_in},
"bronze_punct": {"letters": brz_letters, "records_kept": brz_stats.records_kept,
"records_in": brz_stats.records_in},
"skipped": dict(skipped),
}
(out.parent / "provenance.json").write_text(json.dumps(prov, indent=2))
shutil.rmtree(tmp)
print(f"v1_punct letters: {lit_letters/1e9:.3f}B kept: {lit_stats.records_kept}/{lit_stats.records_in}")
print(f"bronze_punct letters: {brz_letters/1e9:.3f}B kept: {brz_stats.records_kept}/{brz_stats.records_in}")
print(f"skipped (non-training tiers): {dict(skipped)}")
if __name__ == "__main__":
main()
|