Add preview_subset.py (used for josh-hud/code-pii-preview)

preview_subset.py

@@ -0,0 +1,75 @@
+#!/usr/bin/env python3
+"""Build a seeded uniform-random subset of a generated JSONL dataset directly as Parquet.
+
+Every record is kept with probability --frac, decided by a hash of (seed, record id), so the
+sample is reproducible, independent of file order, and stratification follows automatically
+(language / label / positive-negative mix match the parent in expectation).
+
+Usage: python preview_subset.py --src /tmp/poc/v2 --out /tmp/poc/preview_parquet --frac 0.36 --seed 42 --group 9 --workers 4
+"""
+import argparse, collections, glob, hashlib, json, os
+from multiprocessing import Pool
+import pyarrow as pa, pyarrow.parquet as pq
+from to_parquet import SCHEMA
+
+def keep(rid, seed, frac):
+    h = hashlib.blake2b(f"{seed}:{rid}".encode(), digest_size=8).digest()
+    return int.from_bytes(h, "big") / 2 ** 64 < frac
+
+def work(args):
+    srcs, dst, seed, frac = args
+    C = collections.Counter
+    st = {"files": 0, "injected": 0, "spans": 0, "negs": 0, "label": C(), "lang": C(), "value_type": C(), "context": C(), "jsonl_bytes": 0}
+    w = pq.ParquetWriter(dst + ".tmp", SCHEMA, compression="zstd"); buf = []
+    def flush():
+        if buf:
+            w.write_table(pa.Table.from_pylist(buf, schema=SCHEMA)); buf.clear()
+    for src in srcs:
+        with open(src, encoding="utf-8") as fh:
+            for line in fh:
+                r = json.loads(line)
+                if not keep(r["id"], seed, frac):
+                    continue
+                for s in r["hard_negatives"]:
+                    s.setdefault("context", None)
+                buf.append(r); st["files"] += 1; st["jsonl_bytes"] += len(line)
+                st["injected"] += bool(r["spans"]); st["spans"] += len(r["spans"]); st["negs"] += len(r["hard_negatives"])
+                st["lang"][r["language"]] += 1
+                for s in r["spans"]:
+                    st["label"][s["label"]] += 1; st["value_type"][s["value_type"]] += 1; st["context"][s["context"]] += 1
+                if len(buf) >= 10000:
+                    flush()
+    flush(); w.close(); os.replace(dst + ".tmp", dst)
+    st["parquet_bytes"] = os.path.getsize(dst); st["file"] = os.path.basename(dst)
+    print(st["file"], st["files"], "records", f"{st['parquet_bytes']/1e6:.0f} MB", flush=True)
+    return st
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--src", required=True); ap.add_argument("--out", required=True)
+    ap.add_argument("--frac", type=float, default=0.36); ap.add_argument("--seed", type=int, default=42)
+    ap.add_argument("--group", type=int, default=9, help="JSONL parts per output parquet file")
+    ap.add_argument("--workers", type=int, default=4)
+    a = ap.parse_args()
+    os.makedirs(a.out, exist_ok=True)
+    parts = sorted(glob.glob(os.path.join(a.src, "part-*.jsonl")))
+    jobs = []
+    for i in range(0, len(parts), a.group):
+        chunk = parts[i:i + a.group]
+        s, e = os.path.basename(chunk[0])[5:10], os.path.basename(chunk[-1])[5:10]
+        jobs.append((chunk, os.path.join(a.out, f"train-{s}-{e}.parquet"), a.seed, a.frac))
+    with Pool(a.workers) as pool:
+        stats = pool.map(work, jobs)
+    C = collections.Counter
+    agg = {"frac": a.frac, "seed": a.seed, "source_parts": len(parts), "files": 0, "injected": 0, "spans": 0, "negs": 0, "jsonl_bytes": 0, "parquet_bytes": 0,
+           "label": C(), "lang": C(), "value_type": C(), "context": C(), "per_file": []}
+    for st in stats:
+        for k in ("files", "injected", "spans", "negs", "jsonl_bytes", "parquet_bytes"): agg[k] += st[k]
+        for k in ("label", "lang", "value_type", "context"): agg[k].update(st[k])
+        agg["per_file"].append({"file": st["file"], "records": st["files"], "bytes": st["parquet_bytes"]})
+    for k in ("label", "lang", "value_type", "context"): agg[k] = dict(agg[k].most_common())
+    json.dump(agg, open(os.path.join(a.out, "preview_stats.json"), "w"), indent=1)
+    print(json.dumps({k: agg[k] for k in ("files", "injected", "spans", "negs", "jsonl_bytes", "parquet_bytes")}, indent=1))
+
+if __name__ == "__main__":
+    main()