File size: 3,930 Bytes
3111a31
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Algebra version of LFM3's .empryo/tmp/prep_attack_inputs.py (math part only; book inputs are copied unchanged).

Same scheme: per level, kind-proportional stratified sample (50 test / 10 dev per level, seeds 100+lvl / 200+lvl),
translate = first 20 test rows of each level, retain = seeded shuffle per (topic, level) of the non-target topics.
File names keep the LFM3 'math_calc_*' names so the audit code needs no path change; topic is arch.MATH_TARGET=algebra.
"""
import collections
import datetime as dt
import hashlib
import json
import random
from pathlib import Path

D = Path("G:/lexsi/lexsi-unlearning/data")  # frozen data (read-only)
O = Path("attack_data/_inputs")
TARGET = "algebra"
RETAIN = ("arithmetic", "precalculus", "calculus")
RETAIN_PER_LEVEL = {1: 7, 2: 7, 3: 6}  # 20 per topic, 60 total (LFM3: 2 topics x 3 levels x 10)


def rd(p):
    return [json.loads(l) for l in open(D / p, encoding="utf-8") if l.strip()]


def strat(rows, n, seed):  # verbatim from the LFM3 prep script
    by = collections.defaultdict(list)
    for r in rows:
        by[r["kind"]].append(r)
    rng = random.Random(seed)
    out = []
    total = len(rows)
    alloc = {k: max(1, round(n * len(v) / total)) for k, v in by.items()}
    while sum(alloc.values()) > n:
        alloc[max(alloc, key=alloc.get)] -= 1
    while sum(alloc.values()) < n:
        alloc[min(alloc, key=alloc.get)] += 1
    for k, v in sorted(by.items()):
        v = v[:]
        rng.shuffle(v)
        out += v[:alloc[k]]
    rng.shuffle(out)
    return out


def w(name, rows):
    with open(O / name, "w", encoding="utf-8", newline="\n") as f:
        for r in rows:
            f.write(json.dumps(r, ensure_ascii=False) + "\n")
    return len(rows)


c = {}
test, dev = [], []
for lvl in (1, 2, 3):
    test += strat(rd(f"{TARGET}/level{lvl}/test.jsonl"), 50, 100 + lvl)
    dev += strat(rd(f"{TARGET}/level{lvl}/dev.jsonl"), 10, 200 + lvl)
c["math_calc_test"] = w("math_calc_test.jsonl", test)
c["math_calc_dev"] = w("math_calc_dev.jsonl", dev)
tr = [r for lvl in (1, 2, 3) for r in [x for x in test if x["level"] == lvl][:20]]
c["math_calc_test_translate"] = w("math_calc_test_translate.jsonl", tr)
ret = []
for topic in RETAIN:
    for lvl in (1, 2, 3):
        rows = rd(f"{topic}/level{lvl}/test.jsonl")
        random.Random(300 + lvl).shuffle(rows)
        ret += rows[:RETAIN_PER_LEVEL[lvl]]
assert not any(r["topic"] == TARGET for r in ret)
c["math_retain_test"] = w("math_retain_test.jsonl", ret)


def sh(name):
    return hashlib.sha256((O / name).read_bytes()).hexdigest()


def cnt(rows):
    return {f"{t}/L{l}/{k}": n for (t, l, k), n in sorted(collections.Counter((r["topic"], r["level"], r["kind"]) for r in rows).items())}


man = {
    "topic": TARGET, "retain_topics": list(RETAIN),
    "note": "file names keep the LFM3 'math_calc_*' names; contents are ALGEBRA items (arch.MATH_TARGET). "
            "math_retain_test has no algebra rows.",
    "source": str(D).replace("\\", "/") + "/<topic>/level{1,2,3}/{test,dev}.jsonl",
    "scheme": {"math_calc_test": "per level 50, kind-proportional stratified (strat), seed 100+level",
               "math_calc_dev": "per level 10 from dev.jsonl, strat, seed 200+level",
               "math_calc_test_translate": "first 20 rows of each level of math_calc_test",
               "math_retain_test": "per topic in %s, level 1/2/3 -> %s rows after random.Random(300+level).shuffle" % (list(RETAIN), RETAIN_PER_LEVEL)},
    "counts": c,
    "breakdown": {k: cnt(v) for k, v in (("math_calc_test", test), ("math_calc_dev", dev), ("math_calc_test_translate", tr), ("math_retain_test", ret))},
    "sha256": {f"{k}.jsonl": sh(f"{k}.jsonl") for k in c},
    "built": dt.datetime.now(dt.timezone.utc).strftime("%Y-%m-%d %H:%M:%S UTC"),
}
json.dump(man, open(O / "manifest_math.json", "w", encoding="utf-8"), indent=2)
print(json.dumps(man["counts"]), json.dumps(man["breakdown"], indent=1))