File size: 5,227 Bytes
7ed86c3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Stage 7b: verify the documentary-clean corpus before training on it.

  1. census    stream train.jsonl.zst; FAIL if any record has source in
               {ddbdp, dclp, phi} or tier == inscriptions
  2. cleanness sample 1/SAMPLE_EVERY of emitted records; recompute their
               skeleton exact/bag/8-gram hashes and look them up in the
               stage-4b documentary index -> expect ZERO hits (proves the
               excision + re-stitching left no documentary trace)
  3. index     sanity: sample documentary records from work/sentences and
               confirm their keys DO hit the index (the net has no holes)

Exits non-zero on any failure.
"""
import glob
import json
import os
import random
import subprocess
import sys

import numpy as np
import pyarrow.parquet as pq

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from common import h64, skeleton, sentence_spans, NGRAM

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
OUT = os.path.join(ROOT, "work", "doc_clean")
OUTDIR = os.path.expandvars(os.environ.get("DOC_OUTDIR",
                        "$STOICHEIA_DATA"))
PAPYRI_SOURCES = {"ddbdp", "dclp"}
INSCR_SYNTH = {"synthetic", "synthetic_2"}
SAMPLE_EVERY = 25


def record_keys(text):
    """All contamination keys of one text: per-sentence exact+bag, stream 8-grams."""
    keys = []
    words_all = []
    for s, e in sentence_spans(text):
        sk = skeleton(text[s:e])
        w = sk.split()
        if not w:
            continue
        keys.append(h64(sk))
        keys.append(h64(" ".join(sorted(w))))
        words_all.extend(w)
    for i in range(len(words_all) - NGRAM + 1):
        keys.append(h64(" ".join(words_all[i:i + NGRAM])))
    return keys


def main():
    import orjson
    fails = []
    keys_idx = np.load(os.path.join(OUT, "index_keys.npy"), mmap_mode="r")

    def hits(qkeys):
        if not qkeys:
            return 0
        q = np.array(qkeys, dtype=np.uint64)
        pos = np.searchsorted(keys_idx, q)
        pos[pos >= len(keys_idx)] = len(keys_idx) - 1
        return int((keys_idx[pos] == q).sum())

    # ---- 1 + 2: census and cleanness over the emitted corpus ----
    proc = subprocess.Popen(["zstdcat", os.path.join(OUTDIR, "train.jsonl.zst")],
                            stdout=subprocess.PIPE)
    n = 0
    census = {}
    bad_source = 0
    sampled = checked_keys = hit_keys = 0
    hit_examples = []
    for line in proc.stdout:
        r = orjson.loads(line)
        n += 1
        census[(r["tier"], r["source"])] = census.get((r["tier"], r["source"]), 0) + 1
        if r["source"] in PAPYRI_SOURCES or r["source"] == "phi" \
                or r["tier"] == "inscriptions":
            bad_source += 1
        if n % SAMPLE_EVERY == 0:
            ks = record_keys(r["text"])
            h = hits(ks)
            sampled += 1
            checked_keys += len(ks)
            hit_keys += h
            if h and len(hit_examples) < 5:
                hit_examples.append(r["id"])
    if proc.wait() != 0:
        raise RuntimeError("zstdcat failed")
    if bad_source:
        fails.append("census: %d documentary records in output" % bad_source)
    if hit_keys:
        fails.append("cleanness: %d/%d sampled keys hit the doc index (e.g. %s)"
                     % (hit_keys, checked_keys, hit_examples))
    print("records: %d  sampled: %d  keys checked: %d  doc-index hits: %d"
          % (n, sampled, checked_keys, hit_keys))
    by_tier = {}
    for (t, s), c in census.items():
        by_tier[t] = by_tier.get(t, 0) + c
    print("by tier:", json.dumps(by_tier))

    # ---- 3: the index actually contains documentary text ----
    rng = random.Random(0)
    pos_checked = pos_hit = 0
    for tier, want_src in (("inscriptions", None), ("pristine", PAPYRI_SOURCES)):
        shards = sorted(glob.glob(os.path.join(ROOT, "work", "sentences",
                                               tier, "shard_*.parquet")))
        for p in rng.sample(shards, min(3, len(shards))):
            t = pq.read_table(p, columns=["rid", "source", "text"])
            rows = [(r, s, x) for r, s, x in zip(t["rid"].to_pylist(),
                                                 t["source"].to_pylist(),
                                                 t["text"].to_pylist())
                    if (want_src is None and (":" not in r or
                        r.split(":", 1)[1] not in INSCR_SYNTH))
                    or (want_src is not None and s in want_src)]
            for r, s, x in rng.sample(rows, min(50, len(rows))):
                ks = record_keys(x)
                if not ks:
                    continue
                pos_checked += 1
                if hits(ks):
                    pos_hit += 1
    print("index sanity: %d/%d documentary samples hit the index"
          % (pos_hit, pos_checked))
    if pos_checked and pos_hit < pos_checked * 0.98:
        fails.append("index sanity: only %d/%d documentary samples found in index"
                     % (pos_hit, pos_checked))

    if fails:
        print("\nVERIFY FAILED:")
        for f in fails:
            print("  - " + f)
        sys.exit(1)
    print("\nVERIFY OK")


if __name__ == "__main__":
    main()