File size: 13,750 Bytes
1a4b406
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
"""SPROG-9M — standalone inference.

A 9.37M-parameter from-scratch MLX seq2seq that solves grade-school math word
problems WITHOUT an LLM at inference. It abstracts the numbers in a question to
slots ([N0], [N1], ...) and predicts a postfix PROGRAM over those slots; the
program is executed symbolically. Self-consistency (96 temperature samples) plus
a free symbolic verifier (0 trainable params) select the final answer.

Usage:
    pip install mlx numpy huggingface_hub
    huggingface-cli download codelion/sprog-9m --local-dir ./sprog-9m
    python sprog-9m/inference.py --question "A baker had 24 muffins, sold 3/4, then baked 10 more. How many now?"

Python:
    from huggingface_hub import snapshot_download
    from pathlib import Path
    import sys
    p = snapshot_download("codelion/sprog-9m"); sys.path.insert(0, p)
    from inference import load_model, solve
    model, stoi, cfg = load_model(Path(p))
    print(solve(model, stoi, "If 3 pens cost $6, how much do 5 pens cost?"))
"""
from __future__ import annotations

import argparse
import json
import re
from collections import defaultdict
from pathlib import Path

import mlx.core as mx
import mlx.nn as nn
import numpy as np
from mlx.utils import tree_unflatten

# ---------------------------------------------------------------------------
# 1. number-slot tokenization (numbers -> [Ni]; spelled numbers -> digits)
# ---------------------------------------------------------------------------
_UNITS = {"zero": 0, "one": 1, "two": 2, "three": 3, "four": 4, "five": 5,
          "six": 6, "seven": 7, "eight": 8, "nine": 9, "ten": 10, "eleven": 11,
          "twelve": 12, "thirteen": 13, "fourteen": 14, "fifteen": 15,
          "sixteen": 16, "seventeen": 17, "eighteen": 18, "nineteen": 19}
_TENS = {"twenty": 20, "thirty": 30, "forty": 40, "fifty": 50, "sixty": 60,
         "seventy": 70, "eighty": 80, "ninety": 90}
_SCALES = {"hundred": 100, "thousand": 1000, "million": 1_000_000}

OPS = {"+", "-", "*", "/", "**"}
PAD, BOS, EOS, UNK = "<pad>", "<bos>", "<eos>", "<unk>"
SPECIAL = [PAD, BOS, EOS, UNK]
MAX_SLOTS = 20
SLOT_TOKENS = [f"[N{i}]" for i in range(MAX_SLOTS)]
CONST_VALUES = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 15, 16, 20, 24, 25,
                30, 50, 52, 60, 100, 365, 1000,
                0.01, 0.05, 0.1, 0.125, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45,
                0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95,
                1.2, 1.25, 1.5, 2.5]
CONST_TOKENS = [f"[C{c}]" for c in CONST_VALUES]
TGT_VOCAB = SPECIAL + sorted(OPS) + SLOT_TOKENS + CONST_TOKENS
TGT_STOI = {t: i for i, t in enumerate(TGT_VOCAB)}
BOS_ID, EOS_ID = TGT_STOI[BOS], TGT_STOI[EOS]

_NUM = re.compile(r"\d[\d,]*(?:\.\d+)?")
_TOKEN = re.compile(r"\[N\d+\]|[a-z]+|\d+\.?\d*|[^\sa-z\d]")
_WORDTOK = re.compile(r"[A-Za-z]+|\d[\d,]*\.?\d*")


def to_digits(text: str) -> str:
    s = text.replace("-", " ")
    spans, run_start, total, current, have, last_end = [], None, 0, 0, False, 0

    def flush(end):
        nonlocal run_start, total, current, have
        if have and run_start is not None:
            spans.append((run_start, end, str(total + current)))
        run_start, total, current, have = None, 0, 0, False

    for m in _WORDTOK.finditer(s):
        w = m.group().lower()
        if w in _UNITS:
            if not have:
                run_start = m.start()
            current += _UNITS[w]; have = True
        elif w in _TENS:
            if not have:
                run_start = m.start()
            current += _TENS[w]; have = True
        elif w == "hundred" and have:
            current = (current or 1) * 100
        elif w in _SCALES and have:
            total += (current or 1) * _SCALES[w]; current = 0
        elif w == "and" and have:
            pass
        else:
            flush(last_end)
        last_end = m.end()
    flush(last_end)
    if not spans:
        return text
    out, prev = [], 0
    for st, en, rep in spans:
        out.append(s[prev:st]); out.append(rep); prev = en
    out.append(s[prev:])
    return "".join(out)


def slot_encode(text: str):
    """(token_list, slot_values) — numbers replaced by [Ni] in order of appearance."""
    t = to_digits(text).lower()
    values, pieces, prev = [], [], 0
    for m in _NUM.finditer(t):
        values.append(float(m.group().replace(",", "")))
        if len(values) > MAX_SLOTS:
            return None
        pieces.append(t[prev:m.start()]); pieces.append(f" [N{len(values)-1}] ")
        prev = m.end()
    pieces.append(t[prev:])
    return _TOKEN.findall("".join(pieces)), values


# ---------------------------------------------------------------------------
# 2. symbolic execution + verifier (0 trainable params)
# ---------------------------------------------------------------------------
def decode_postfix(seq, values):
    stack = []
    for t in seq:
        if t in OPS:
            if len(stack) < 2:
                return None
            b, a = stack.pop(), stack.pop()
            if t == "/":
                if b == 0:
                    return None
                stack.append(a / b)
            elif t == "**":
                if abs(a) > 1e4 or abs(b) > 6 or (a < 0 and b != int(b)) or (a == 0 and b < 0):
                    return None
                try:
                    r = a ** b
                except (OverflowError, ValueError, ZeroDivisionError):
                    return None
                if isinstance(r, complex):
                    return None
                stack.append(r)
            else:
                stack.append({"+": a + b, "-": a - b, "*": a * b}[t])
        elif t.startswith("[N"):
            i = int(t[2:-1])
            if i >= len(values):
                return None
            stack.append(values[i])
        elif t.startswith("[C"):
            stack.append(float(t[2:-1]))
        else:
            return None
    return stack[0] if len(stack) == 1 else None


def _intermediates(prog, vals):
    s, inter = [], []
    for t in prog:
        if t in OPS:
            if len(s) < 2:
                return None
            b, a = s.pop(), s.pop()
            try:
                if t == "+": r = a + b
                elif t == "-": r = a - b
                elif t == "*": r = a * b
                elif t == "/": r = a / b if b != 0 else None
                elif t == "**": r = a ** b if not (a < 0 and b != int(b)) and not (a == 0 and b < 0) else None
                else: r = None
            except Exception:
                return None
            if r is None or isinstance(r, complex):
                return None
            s.append(r); inter.append(r)
        elif t.startswith("[N"):
            i = int(t[2:-1])
            if i >= len(vals):
                return None
            s.append(vals[i])
        elif t.startswith("[C"):
            s.append(float(t[2:-1]))
        else:
            return None
    return inter


def _plausible(a):
    return a is not None and a > 0 and abs(a - round(a)) < 1e-6


def verify_select(question, candidates, mag_k=100.0):
    """Pick the answer whose best program maximizes the structural score
    (coverage + magnitude-sanity + intermediate-sanity), tie-broken by votes."""
    enc = slot_encode(question)
    vals = enc[1] if enc else []
    navail = max(len(vals), 1)
    nums = [float(x) for x in re.findall(r"\d+\.?\d*", question)]
    maxn = max(nums) if nums else 1.0
    votes, V = defaultdict(float), defaultdict(lambda: -1e9)
    for prog, a in candidates:
        votes[a] += 1.0
        cov = len({t for t in prog if t.startswith("[N")}) / navail
        mag = 1.0 if a <= mag_k * maxn else 0.0
        inter = _intermediates(prog, vals)
        ip = 1.0 if (inter is not None and all(x >= 0 for x in inter)) else 0.0
        ii = 1.0 if (inter is not None and all(abs(x - round(x)) < 1e-6 for x in inter)) else 0.0
        V[a] = max(V[a], cov + mag + ip + ii)
    if not votes:
        return None
    plau = [a for a in votes if _plausible(a)] or list(votes)
    return max(plau, key=lambda a: (V[a], votes[a]))


# ---------------------------------------------------------------------------
# 3. the MLX seq2seq model
# ---------------------------------------------------------------------------
def _pad_mask(pad):
    return mx.where(pad[:, None, None, :], -1e9, 0.0).astype(mx.float32)


def _causal_mask(L):
    return mx.triu(mx.full((L, L), -1e9), k=1)[None, None]


class FFN(nn.Module):
    def __init__(self, d, ff):
        super().__init__(); self.l1 = nn.Linear(d, ff); self.l2 = nn.Linear(ff, d)

    def __call__(self, x):
        return self.l2(nn.gelu(self.l1(x)))


class EncLayer(nn.Module):
    def __init__(self, d, h, ff):
        super().__init__()
        self.n1 = nn.LayerNorm(d); self.attn = nn.MultiHeadAttention(d, h)
        self.n2 = nn.LayerNorm(d); self.ffn = FFN(d, ff)

    def __call__(self, x, m):
        h = self.n1(x); x = x + self.attn(h, h, h, m)
        h = self.n2(x); return x + self.ffn(h)


class DecLayer(nn.Module):
    def __init__(self, d, h, ff):
        super().__init__()
        self.n1 = nn.LayerNorm(d); self.sa = nn.MultiHeadAttention(d, h)
        self.n2 = nn.LayerNorm(d); self.ca = nn.MultiHeadAttention(d, h)
        self.n3 = nn.LayerNorm(d); self.ffn = FFN(d, ff)

    def __call__(self, x, mem, tm, mm):
        h = self.n1(x); x = x + self.sa(h, h, h, tm)
        h = self.n2(x); x = x + self.ca(h, mem, mem, mm)
        h = self.n3(x); return x + self.ffn(h)


class Seq2Seq(nn.Module):
    def __init__(self, src_v, tgt_v, d=304, h=4, ne=4, nd=4, ff=608,
                 max_src=220, max_tgt=64):
        super().__init__()
        self.src_emb = nn.Embedding(src_v, d); self.tgt_emb = nn.Embedding(tgt_v, d)
        self.src_pos = nn.Embedding(max_src, d); self.tgt_pos = nn.Embedding(max_tgt, d)
        self.enc = [EncLayer(d, h, ff) for _ in range(ne)]
        self.dec = [DecLayer(d, h, ff) for _ in range(nd)]
        self.norm = nn.LayerNorm(d); self.out = nn.Linear(d, tgt_v)

    def encode(self, src, src_pad):
        L = src.shape[1]
        x = self.src_emb(src) + self.src_pos(mx.arange(L))
        m = _pad_mask(src_pad)
        for layer in self.enc:
            x = layer(x, m)
        return x, m

    def decode(self, tgt, mem, mm, tgt_pad):
        L = tgt.shape[1]
        x = self.tgt_emb(tgt) + self.tgt_pos(mx.arange(L))
        tm = _causal_mask(L) + _pad_mask(tgt_pad)
        for layer in self.dec:
            x = layer(x, mem, tm, mm)
        return self.out(self.norm(x))


def _sample_batch(model, mem, mm, S, T, maxlen=48):
    memS = mx.broadcast_to(mem, (S, mem.shape[1], mem.shape[2]))
    mmS = mx.broadcast_to(mm, (S, mm.shape[1], mm.shape[2], mm.shape[3]))
    seqs = mx.full((S, 1), BOS_ID, dtype=mx.int32)
    for _ in range(maxlen):
        logits = model.decode(seqs, memS, mmS,
                              mx.zeros((S, seqs.shape[1]), dtype=mx.bool_))
        nxt = mx.random.categorical(logits[:, -1, :] / T)[:, None]
        seqs = mx.concatenate([seqs, nxt.astype(mx.int32)], axis=1)
    seqs, out = np.array(seqs), []
    for row in seqs:
        toks = []
        for t in row[1:]:
            if t == EOS_ID:
                break
            toks.append(TGT_VOCAB[int(t)])
        out.append(toks)
    return out


# ---------------------------------------------------------------------------
# 4. load + solve
# ---------------------------------------------------------------------------
def load_model(model_dir: Path):
    model_dir = Path(model_dir)
    cfg = json.load(open(model_dir / "config.json"))
    src_vocab = json.load(open(model_dir / "src_vocab.json"))
    stoi = {t: i for i, t in enumerate(src_vocab)}
    model = Seq2Seq(len(src_vocab), len(TGT_VOCAB), d=cfg["d"], h=cfg["n_heads"],
                    ne=cfg["n_layers"], nd=cfg["n_layers"], ff=cfg["ff"],
                    max_src=cfg["max_src"], max_tgt=cfg["max_tgt"])
    w = np.load(model_dir / "model.npz")
    model.update(tree_unflatten([(k, mx.array(w[k])) for k in w.files]))
    model.eval()
    return model, stoi, cfg


def solve(model, stoi, question, n_samples=96, temp=0.9, seed=0, return_details=False):
    """Solve one question. Returns the answer (float) or None."""
    mx.random.seed(seed)
    enc = slot_encode(question)
    if enc is None:
        return None
    toks, vals = enc
    S = mx.array([[stoi.get(t, stoi[UNK]) for t in toks]])
    mem, mm = model.encode(S, S == 0)
    cands = []
    for prog in _sample_batch(model, mem, mm, n_samples, temp):
        a = decode_postfix(prog, vals)
        if a is None or isinstance(a, complex):
            continue
        cands.append((prog, round(float(a), 4)))
    if not cands:
        return None
    answer = verify_select(question, cands)
    if return_details:
        votes = defaultdict(int)
        for _, a in cands:
            votes[a] += 1
        return {"answer": answer, "n_candidates": len(cands),
                "n_distinct": len(votes), "top_votes": sorted(votes.items(),
                key=lambda x: -x[1])[:5]}
    return answer


def main():
    ap = argparse.ArgumentParser(description="SPROG-9M: LLM-free GSM8K solver")
    ap.add_argument("--question", "-q", required=True)
    ap.add_argument("--model-dir", default=str(Path(__file__).parent))
    ap.add_argument("--samples", type=int, default=96)
    ap.add_argument("--temp", type=float, default=0.9)
    ap.add_argument("--details", action="store_true")
    a = ap.parse_args()
    model, stoi, cfg = load_model(Path(a.model_dir))
    r = solve(model, stoi, a.question, a.samples, a.temp, return_details=a.details)
    if a.details:
        print(json.dumps(r, indent=2))
    else:
        print(f"Answer: {r}")


if __name__ == "__main__":
    main()