File size: 10,553 Bytes
96ac6f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7debf49
 
 
96ac6f7
 
 
 
 
 
 
 
 
7debf49
 
 
 
 
 
 
 
 
96ac6f7
 
 
 
 
 
 
 
 
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
"""Babak 1b: parallel math word problems with code-verified answers, then solved per language/condition.

1) Beam writes an Indonesian problem + Python solver (effort low); solver executed; keep if it reproduces the answer.

2) Beam translates the full prompt into 8 languages; keep only if every number survives the translation.

3) Beam solves each version (conditions C0/C2) -> data/math/runs_<effort>.jsonl with reasoning-language profile.

Usage: python mathset.py gen --n 70 ; python mathset.py translate ; python mathset.py solve --effort low"""
import argparse, json, math, os, random, re, subprocess, sys, tempfile
from beam import chat, run_many, spent
from lid import profile

OUT = "data/math"
LANGS = {"ind_Latn": "Indonesian", "eng_Latn": "English", "jav_Latn": "Javanese", "sun_Latn": "Sundanese",
         "zsm_Latn": "Malay", "tgl_Latn": "Tagalog", "ceb_Latn": "Cebuano", "vie_Latn": "Vietnamese", "tha_Thai": "Thai"}
TOPICS = ["belanja dan diskon di pasar", "gaji, THR dan potongan dengan tarif yang disebut di soal",
          "cicilan dan bunga flat", "zakat dengan nisab dan kadar yang disebut di soal", "ongkos kirim antarpulau",
          "jarak, kecepatan dan zona waktu WIB/WITA/WIT", "hasil panen sawah per hektar", "modal dan untung warung",
          "takaran resep untuk hajatan", "arisan dan koperasi", "luas dan keliling rumah atau kolam",
          "peluang dan kombinatorika", "barisan dan pola tabungan", "rata-rata dan median nilai rapor",
          "FPB dan KPK jadwal", "sistem persamaan dari cerita umur"]
ANSWER_LINE = ("Tulis jawaban akhir sebagai angka biasa (tanpa pemisah ribuan, titik untuk desimal) "
               "di dalam \\boxed{}.")
GEN = """Buat SATU soal cerita matematika orisinal dalam bahasa Indonesia yang alami. Topik: {topic}. \

Tingkat kesulitan {d}/5 (1=SD, 3=SMP, 5=olimpiade).

Aturan: semua angka/aturan yang dibutuhkan tertulis di soal; jawaban akhir tepat SATU bilangan (sebutkan pembulatan \

bila perlu); soal punya tepat satu jawaban; tulis angka dengan digit. "solver" = kode Python murni (boleh import math, \

fractions, itertools saja) yang print() HANYA jawaban akhir sebagai angka biasa.

Keluarkan HANYA JSON: {{"soal": "...", "jawaban": <angka>, "solver": "..."}}"""
TRANSLATE = """Translate the following Indonesian math word problem into natural, fluent {name}. Keep every number \

exactly as written with Western digits (do not convert numbers to words, do not change units or currency amounts). \

Translate the final instruction sentence too. Output ONLY the translation.



{text}"""
SYS = {"ind_Latn": "Berpikir dan bernalarlah langkah demi langkah dalam bahasa Indonesia."}  # others from atlas templates

BANNED = re.compile(r"\b(import\s+(os|sys|subprocess|socket|shutil|urllib|pathlib|ctypes|requests)|from\s+"
                    r"(os|sys|subprocess|socket|shutil|urllib|pathlib|ctypes|requests)|open\s*\(|exec\s*\(|eval\s*\(|__import__)")


def run_solver(code, timeout=10):
    # ponytail: denylist + timeout + temp dir, not a real sandbox; fine for our own generated snippets
    if BANNED.search(code):
        return None
    with tempfile.TemporaryDirectory() as d:
        p = os.path.join(d, "s.py"); open(p, "w", encoding="utf-8").write(code)
        try:
            r = subprocess.run([sys.executable, "-I", p], cwd=d, capture_output=True, text=True, timeout=timeout)
        except subprocess.TimeoutExpired:
            return None
    lines = r.stdout.strip().splitlines()
    return lines[-1].strip() if r.returncode == 0 and lines else None


def nums(s):
    """Float readings of the last number in s; ambiguous '1.250' yields both 1250 and 1.25."""
    if s is None:
        return set()
    s = re.sub(r"(?i)rp\.?|%|\s|\\[,!]|\$|\\text\{[^}]*\}", "", str(s))
    fr = re.search(r"(-?)\\d?frac\{(-?\d+)\}\{(\d+)\}", s)
    if fr and int(fr.group(3)):
        return {(-1 if fr.group(1) else 1) * int(fr.group(2)) / int(fr.group(3))}
    m = re.findall(r"-?\d[\d.,]*", s)
    if not m:
        return set()
    t = m[-1].rstrip(".,"); c = set()
    if "," in t and "." in t:
        c.add(t.replace(".", "").replace(",", ".") if t.rfind(",") > t.rfind(".") else t.replace(",", ""))
    elif t.count(".") > 1 or t.count(",") > 1:
        c.add(t.replace(".", "").replace(",", ""))
    elif "." in t or "," in t:
        a, b = re.split(r"[.,]", t); c.add(f"{a}.{b}")
        if len(b) == 3:
            c.add(a + b)
    else:
        c.add(t)
    out = set()
    for x in c:
        try:
            out.add(float(x))
        except ValueError:
            pass
    return out


def same(text, answer):
    return any(math.isclose(v, float(answer), rel_tol=1e-4, abs_tol=0.011) for v in nums(text))


def boxed(text):
    i = (text or "").rfind("\\boxed{")
    if i < 0:
        return None
    i += 7; depth = 1; j = i
    while j < len(text) and depth:
        depth += {"{": 1, "}": -1}.get(text[j], 0); j += 1
    return text[i:j - 1] if depth == 0 else None


def digit_multiset(text):
    """Numbers in text, separator-insensitive: thousands marks dropped, decimal comma == decimal point."""
    return sorted(re.sub(r"[.,](?=\d{3}\b)", "", x).replace(",", ".")
                  for x in re.findall(r"\d[\d.,]*\d|\d", text))


def cmd_gen(a):
    os.makedirs(OUT, exist_ok=True)

    def one(i):
        rng = random.Random(i)
        topic, d = rng.choice(TOPICS), rng.choice([1, 2, 3, 3, 4, 4, 5])
        r = chat([{"role": "user", "content": GEN.format(topic=topic, d=d)}], "low", 8192, tag=f"gen{i}")
        m = re.search(r"\{.*\}", r["content"], re.S)
        if not m:
            return {"id": i, "ok": False, "why": "unparseable"}
        try:
            spec = json.loads(m.group(0)); ans = float(spec["jawaban"])
        except Exception:
            return {"id": i, "ok": False, "why": "unparseable"}
        out = run_solver(spec["solver"])
        ok = out is not None and same(out, ans)
        return {"id": i, "ok": ok, "why": "ok" if ok else "solver_mismatch", "topic": topic, "difficulty": d,
                "question": spec["soal"].strip(), "answer": ans, "solver": spec["solver"], "solver_output": out}

    rows = [res for _, res in run_many(one, list(range(a.n)), 12, "gen") if not isinstance(res, Exception)]
    json.dump(sorted(rows, key=lambda x: x["id"]), open(f"{OUT}/problems.json", "w", encoding="utf-8"),
              ensure_ascii=False, indent=1)
    print(sum(r["ok"] for r in rows), "verified of", len(rows))


def cmd_translate(a):
    probs = [p for p in json.load(open(f"{OUT}/problems.json", encoding="utf-8")) if p["ok"]]
    jobs = [(p, c) for p in probs for c in LANGS if c != "ind_Latn"]

    def one(job):
        p, c = job
        src = f"{p['question']}\n\n{ANSWER_LINE}"
        r = chat([{"role": "user", "content": TRANSLATE.format(name=LANGS[c], text=src)}], "low", 4096,
                 temperature=0.2, tag="tr")
        t = r["content"].strip()
        return {"id": p["id"], "lang": c, "text": t, "numbers_ok": digit_multiset(src) == digit_multiset(t),
                "lid": profile(t, c)}

    par = {p["id"]: {"ind_Latn": {"text": f"{p['question']}\n\n{ANSWER_LINE}", "numbers_ok": True}} for p in probs}
    for _, res in run_many(one, jobs, 16, "translate"):
        if not isinstance(res, Exception):
            par[res["id"]][res["lang"]] = res
    json.dump(par, open(f"{OUT}/parallel.json", "w", encoding="utf-8"), ensure_ascii=False, indent=1)
    full = [i for i, v in par.items() if all(c in v and v[c]["numbers_ok"] for c in LANGS)]
    print(f"{len(full)}/{len(par)} problems survive in all {len(LANGS)} languages")


def cmd_solve(a):
    probs = {p["id"]: p for p in json.load(open(f"{OUT}/problems.json", encoding="utf-8")) if p["ok"]}
    par = {int(k): v for k, v in json.load(open(f"{OUT}/parallel.json", encoding="utf-8")).items()}
    tps = json.load(open("data/atlas/templates.json", encoding="utf-8"))
    ids = sorted(i for i, v in par.items() if all(c in v and v[c]["numbers_ok"] for c in LANGS))[:a.n]
    langs = a.langs or list(LANGS)
    jobs = [(i, c, cond) for i in ids for c in langs for cond in a.conds]
    budget = 8192 if a.effort == "low" else 32768

    def one(job):
        i, c, cond = job
        msgs = [{"role": "user", "content": par[i][c]["text"]}]
        if cond == "C2":
            msgs.insert(0, {"role": "system", "content": SYS.get(c) or tps[c]["reason_instruction"]})
        elif cond == "C3":  # the shipped benak_lang wrapper; its primer is a reading item -> tests transfer to math
            from benak_lang import think_in
            msgs = think_in(msgs, c, primer=True)
        r = chat(msgs, a.effort, budget, tag="math")
        b = boxed(r["content"])
        return {"id": i, "lang": c, "cond": cond, "effort": a.effort, "answer": probs[i]["answer"], "boxed": b,
                "correct": b is not None and same(b, probs[i]["answer"]), "finish_reason": r["finish_reason"],
                "usage": r["usage"], "reasoning": r["reasoning"], "content": r["content"],
                "lid": profile(r["reasoning"], c), "lid_answer": profile(r["content"], c),
                "topic": probs[i]["topic"], "difficulty": probs[i]["difficulty"], "route": r.get("route"), "served_model": r.get("model"), "latency_s": r.get("latency_s")}

    rows = [res for _, res in run_many(one, jobs, 16, f"math-{a.effort}") if not isinstance(res, Exception)]
    path = f"{OUT}/runs_{a.effort}.jsonl"
    merged = {}
    if os.path.exists(path):  # merge: a run with other --conds/--langs must not erase earlier rows
        for l in open(path, encoding="utf-8"):
            d = json.loads(l); merged[(d["id"], d["lang"], d["cond"])] = d
    for r in rows:
        merged[(r["id"], r["lang"], r["cond"])] = r
    with open(path, "w", encoding="utf-8") as f:
        f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in merged.values())
    print(f"{len(rows)}/{len(jobs)} solved-rows written; spent ${spent()}")


if __name__ == "__main__":
    ap = argparse.ArgumentParser(); ap.add_argument("cmd", choices=["gen", "translate", "solve"])
    ap.add_argument("--n", type=int, default=70); ap.add_argument("--effort", default="low")
    ap.add_argument("--conds", nargs="+", default=["C0", "C2"]); ap.add_argument("--langs", nargs="+")
    a = ap.parse_args()
    {"gen": cmd_gen, "translate": cmd_translate, "solve": cmd_solve}[a.cmd](a)