pass5 stats script 9f697e6f
Browse files
code/pass5/9f697e6f/abeja_v5_stats.py
ADDED
|
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""第5パス出力(outputs_v5)のジャンル・製造業適合・ホスト分布の集計。
|
| 2 |
+
|
| 3 |
+
各シャードの ``pass5_sha1.parquet``(保持文書の本文 SHA-1)と第2パスの ``features.parquet`` を
|
| 4 |
+
SHA-1 で突き合わせ、ジャンル別件数、製造業・重工の適合(manufacturing_score >= しきい値)、
|
| 5 |
+
文字数を数える。マスク置換された文書は SHA-1 が変わるため未突合として件数を残す。
|
| 6 |
+
"""
|
| 7 |
+
from __future__ import annotations
|
| 8 |
+
import argparse, collections, json, os, time
|
| 9 |
+
from concurrent.futures import ProcessPoolExecutor
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
|
| 12 |
+
BUCKETS = [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 5.0, 10.0]
|
| 13 |
+
CHAR_BUCKETS = [500, 1000, 2000, 4000, 8000, 16000, 32000]
|
| 14 |
+
PPL_BUCKETS = [10000, 20000, 30000, 40000, 60000, 100000]
|
| 15 |
+
HOST_TOPN = 5000
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def _bucket(value: float, edges: list) -> str:
|
| 19 |
+
return next((f"<{e}" for e in edges if value < e), f">={edges[-1]}")
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def _suffix(host: str) -> str:
|
| 23 |
+
"""ホストの種別(co.jp / ne.jp / com など)。日本の属性型 JP ドメインは 2 段で見る。"""
|
| 24 |
+
parts = (host or "").lower().rstrip(".").split(".")
|
| 25 |
+
if len(parts) < 2: return "unknown"
|
| 26 |
+
if parts[-1] == "jp" and len(parts) >= 3 and parts[-2] in {"co", "ne", "or", "ac", "go", "ed", "gr", "lg", "ad"}: return ".".join(parts[-2:])
|
| 27 |
+
return parts[-1]
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
def _fetch(repo: str, path: str, staging: str) -> str:
|
| 31 |
+
from huggingface_hub import hf_hub_download
|
| 32 |
+
for k in range(6):
|
| 33 |
+
try: return hf_hub_download(repo, path, repo_type="dataset", token=os.environ.get("HF_TOKEN"), local_dir=staging)
|
| 34 |
+
except Exception: # noqa: BLE001
|
| 35 |
+
if k == 5: raise
|
| 36 |
+
time.sleep(20 * (k + 1))
|
| 37 |
+
raise RuntimeError("unreachable")
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def one(task: dict) -> dict:
|
| 41 |
+
import pyarrow.parquet as pq
|
| 42 |
+
t0 = time.perf_counter(); sh = task["shard"]; thr = task["threshold"]
|
| 43 |
+
try:
|
| 44 |
+
sha_f = _fetch(task["repo"], task["sha1_path"], task["staging"])
|
| 45 |
+
rows_v5 = 0; remaining: collections.Counter = collections.Counter()
|
| 46 |
+
pf = pq.ParquetFile(sha_f)
|
| 47 |
+
for b in pf.iter_batches(batch_size=65536, columns=["sha1"]):
|
| 48 |
+
vals = b.column("sha1").to_pylist(); rows_v5 += len(vals); remaining.update(v.hex() for v in vals)
|
| 49 |
+
Path(sha_f).unlink(missing_ok=True)
|
| 50 |
+
feat_f = _fetch(task["repo"], task["features_path"], task["staging"])
|
| 51 |
+
genre = collections.Counter(); manu = collections.Counter(); hist = collections.Counter()
|
| 52 |
+
chars = 0; chars_manu = 0; matched = 0; low_quality = 0
|
| 53 |
+
hosts = collections.Counter(); suffixes = collections.Counter(); char_hist = collections.Counter(); ppl_hist = collections.Counter()
|
| 54 |
+
cols = ["content_sha1", "genre", "manufacturing_score", "chars_final", "is_low_quality", "host", "perplexity"]
|
| 55 |
+
for b in pq.ParquetFile(feat_f).iter_batches(batch_size=65536, columns=cols):
|
| 56 |
+
d = b.to_pydict()
|
| 57 |
+
for i in range(b.num_rows):
|
| 58 |
+
key = d["content_sha1"][i]
|
| 59 |
+
if not key or remaining.get(key, 0) <= 0: continue
|
| 60 |
+
remaining[key] -= 1; matched += 1
|
| 61 |
+
g = d["genre"][i] or "unknown"; genre[g] += 1
|
| 62 |
+
c = d["chars_final"][i] or 0; chars += c
|
| 63 |
+
h = (d["host"][i] or "").lower(); hosts[h] += 1; suffixes[_suffix(h)] += 1
|
| 64 |
+
char_hist[_bucket(c, CHAR_BUCKETS)] += 1
|
| 65 |
+
pp = d["perplexity"][i]; ppl_hist["null" if pp is None else _bucket(pp, PPL_BUCKETS)] += 1
|
| 66 |
+
if d["is_low_quality"][i]: low_quality += 1
|
| 67 |
+
m = d["manufacturing_score"][i]
|
| 68 |
+
if m is None: hist["null"] += 1; continue
|
| 69 |
+
hist[next((f"<{b_}" for b_ in BUCKETS if m < b_), ">=10.0")] += 1
|
| 70 |
+
if m >= thr: manu[g] += 1; chars_manu += c
|
| 71 |
+
Path(feat_f).unlink(missing_ok=True)
|
| 72 |
+
return {"shard": sh, "hosts": dict(hosts.most_common(HOST_TOPN)), "suffixes": dict(suffixes),
|
| 73 |
+
"chars_hist": dict(char_hist), "perplexity_hist": dict(ppl_hist), "rows_v5": rows_v5, "matched": matched, "unmatched": rows_v5 - matched, "by_genre": dict(genre),
|
| 74 |
+
"manufacturing_by_genre": dict(manu), "manufacturing_score_hist": dict(hist), "chars": chars,
|
| 75 |
+
"chars_manufacturing": chars_manu, "low_quality": low_quality, "seconds": round(time.perf_counter() - t0, 1)}
|
| 76 |
+
except Exception as error: # noqa: BLE001
|
| 77 |
+
return {"shard": sh, "error": repr(error)}
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def main() -> None:
|
| 81 |
+
p = argparse.ArgumentParser(); p.add_argument("--repo", required=True); p.add_argument("--dir", required=True)
|
| 82 |
+
p.add_argument("--v5-prefix", default="outputs_v5"); p.add_argument("--features-prefix", default="outputs_v2")
|
| 83 |
+
p.add_argument("--threshold", type=float, default=2.0); p.add_argument("--out", type=Path, default=Path("/work/out"))
|
| 84 |
+
p.add_argument("--staging", default="/work/staging"); p.add_argument("--limit-shards", type=int, default=None)
|
| 85 |
+
p.add_argument("--upload-repo", default=None); p.add_argument("--upload-prefix", default="reports/v5_stats")
|
| 86 |
+
p.add_argument("--workers", type=int, default=max(1, int(os.environ.get("CPU_CORES") or os.cpu_count() or 2) - 1))
|
| 87 |
+
a = p.parse_args()
|
| 88 |
+
from huggingface_hub import HfApi
|
| 89 |
+
api = HfApi(token=os.environ.get("HF_TOKEN"))
|
| 90 |
+
shards = sorted({e.path.split("/")[-1] for e in api.list_repo_tree(a.repo, path_in_repo=f"{a.v5_prefix}/{a.dir}", repo_type="dataset", recursive=False)})
|
| 91 |
+
if a.limit_shards: shards = shards[: a.limit_shards]
|
| 92 |
+
tasks = [{"repo": a.repo, "shard": sh, "staging": a.staging, "threshold": a.threshold,
|
| 93 |
+
"sha1_path": f"{a.v5_prefix}/{a.dir}/{sh}/pass5_sha1.parquet",
|
| 94 |
+
"features_path": f"{a.features_prefix}/{a.dir}/{sh}/features.parquet"} for sh in shards]
|
| 95 |
+
tot = collections.Counter(); genre = collections.Counter(); manu = collections.Counter(); hist = collections.Counter(); failed = []; shard_rows = {}
|
| 96 |
+
hosts = collections.Counter(); suffixes = collections.Counter(); char_hist = collections.Counter(); ppl_hist = collections.Counter()
|
| 97 |
+
with ProcessPoolExecutor(max_workers=a.workers) as ex:
|
| 98 |
+
for r in ex.map(one, tasks):
|
| 99 |
+
if "error" in r: failed.append(r["shard"]); print(json.dumps({"event": "error", **r}), flush=True); continue
|
| 100 |
+
for k in ("rows_v5", "matched", "unmatched", "chars", "chars_manufacturing", "low_quality"): tot[k] += r[k]
|
| 101 |
+
genre.update(r["by_genre"]); manu.update(r["manufacturing_by_genre"]); hist.update(r["manufacturing_score_hist"])
|
| 102 |
+
hosts.update(r["hosts"]); suffixes.update(r["suffixes"]); char_hist.update(r["chars_hist"]); ppl_hist.update(r["perplexity_hist"])
|
| 103 |
+
shard_rows[r["shard"]] = r["rows_v5"]; print(json.dumps({"event": "done", "shard": r["shard"], "rows_v5": r["rows_v5"], "seconds": r["seconds"]}), flush=True)
|
| 104 |
+
result = {"dir": a.dir, "shards": len(tasks), "failed": failed, "threshold": a.threshold, "totals": dict(tot),
|
| 105 |
+
"by_genre": dict(genre), "manufacturing_by_genre": dict(manu), "manufacturing_total": sum(manu.values()),
|
| 106 |
+
"manufacturing_score_hist": dict(hist), "suffixes": dict(suffixes), "chars_hist": dict(char_hist),
|
| 107 |
+
"perplexity_hist": dict(ppl_hist), "hosts_top": dict(hosts.most_common(3000)), "hosts_distinct_seen": len(hosts),
|
| 108 |
+
"shard_rows": shard_rows}
|
| 109 |
+
out = a.out / a.dir; out.mkdir(parents=True, exist_ok=True)
|
| 110 |
+
(out / "v5_stats.json").write_text(json.dumps(result, ensure_ascii=False, indent=1) + "\n", encoding="utf-8")
|
| 111 |
+
if a.upload_repo:
|
| 112 |
+
for k in range(6):
|
| 113 |
+
try: api.upload_folder(repo_id=a.upload_repo, repo_type="dataset", folder_path=str(out), path_in_repo=f"{a.upload_prefix}/{a.dir}", commit_message=f"v5 stats {a.dir}"); break
|
| 114 |
+
except Exception: # noqa: BLE001
|
| 115 |
+
if k == 5: raise
|
| 116 |
+
time.sleep(60 * (k + 1))
|
| 117 |
+
print(json.dumps({"event": "finish", "dir": a.dir, "totals": dict(tot), "manufacturing_total": sum(manu.values()), "failed": failed}, ensure_ascii=False), flush=True)
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
if __name__ == "__main__":
|
| 121 |
+
main()
|