File size: 27,003 Bytes
b65604f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
ROSA Plus v2 (character-level)
--------------------------------
A boundary-aware Suffix Automaton (SAM) with a probabilistic fallback LM
(Witten–Bell interpolation along suffix links). This module exposes a single
high-level class `ROSAPlus` and supporting classes that let you:

- Create a model object
- Train on individual samples (examples) *without* crossing boundaries
- Build the fallback LM (from accumulated examples) and keep the SAM frozen at inference
- Save/load a JSON model
- Provide a pre-existing context and:
  * generate text (string) with sampling controls, or
  * query the probability of the next character (or the full next-char distribution)

This is a modular split of the original script into a library you can import.
The companion CLI wrapper `rosa_plus_train.py` maintains the original flags.
"""
from __future__ import annotations
from collections import deque
from typing import List, Optional, Dict, Tuple
from collections import defaultdict, Counter
import json as _stdlib_json
import math
import random

try:
    import orjson as _json
    _dumps = _json.dumps  # returns bytes
    _loads = _json.loads
except Exception:  # pragma: no cover
    # Fallback to stdlib json (slower, but keeps things working)
    _json = _stdlib_json
    _dumps = lambda obj: _stdlib_json.dumps(obj).encode("utf-8")
    _loads = lambda b: _stdlib_json.loads(b.decode("utf-8"))

from tqdm import tqdm

__all__ = [
    "load_examples_from_file",
    "ROSACharPredictor",
    "ROSAFallbackLM",
    "ROSAPlus",
    "ROSAGRUAdapter",
]


# =========================
# 1) Load & split examples
# =========================

def load_examples_from_file(
    path: str,
    delimiter: str = "<|ENDOFTEXT|>",
    strip_each: bool = True,
    use_eot: bool = True,
    eot_char: str = "\u0004",  # single-char EOT (U+0004 by default)
) -> List[str]:
    """Reads a file, splits by `delimiter`, optional strip, drops empties.
    If `use_eot`, appends a single-character EOT to each example.
    """
    if not isinstance(eot_char, str) or len(eot_char) != 1:
        raise ValueError("eot_char must be a single character.")
    with open(path, "r", encoding="utf-8") as f:
        raw = f.read()
    parts = raw.split(delimiter)
    if strip_each:
        parts = [p.strip() for p in parts]
    parts = [p for p in parts if p]

    if use_eot:
        parts = [p + eot_char for p in parts]

    return parts[:100000]


# =========================
# 2) Boundary-aware SAM
# =========================

class ROSACharPredictor:
    """
    Online/streaming SAM builder (ROSA core), boundary-aware.
    Arrays:
      - b: transitions (list[dict(char->next_state)])
      - c: suffix links (list[int])
      - d: max length (list[int])
      - e: rightmost previous end position per state (list[int])
      - g: last (active) state (int)
    Additionally:
      - text: list[str] training characters in sequence
      - boundary_after[i] = True if an example boundary occurs immediately AFTER position i
    """
    def __init__(self):
        self.b: List[Dict[str, int]] = [{}]  # transitions
        self.c: List[int] = [-1]             # suffix links
        self.d: List[int] = [0]              # max length
        self.e: List[int] = [-1]             # rightmost previous indices
        self.g: int = 0                      # last state
        self.text: List[str] = []            # stores training text characters
        self.boundary_after: List[bool] = [] # boundary flags per char index

    def feed(self, ch: str) -> None:
        """Extend SAM with `ch` (training step). Deterministic prediction is computed at inference."""
        i = len(self.text)
        self.text.append(ch)
        if len(self.boundary_after) < len(self.text):
            self.boundary_after.append(False)

        b, c, d, e, g = self.b, self.c, self.d, self.e, self.g

        # --- SAM extend ---
        r = len(b)
        b.append({})
        c.append(0)
        d.append(d[g] + 1)
        e.append(-1)
        p = g
        while p != -1 and ch not in b[p]:
            b[p][ch] = r
            p = c[p]
        if p == -1:
            c[r] = 0
        else:
            q = b[p][ch]
            if d[p] + 1 == d[q]:
                c[r] = q
            else:
                # clone q -> u
                u = len(b)
                b.append(b[q].copy())
                c.append(c[q])
                d.append(d[p] + 1)
                e.append(e[q])
                while p != -1 and b[p].get(ch) == q:
                    b[p][ch] = u
                    p = c[p]
                c[q] = c[r] = u
        self.g = r

        # --- update rightmost indices ---
        v = self.g
        while v != -1 and self.e[v] < i:
            self.e[v] = i
            v = self.c[v]

    def mark_boundary(self) -> None:
        """Call after finishing an example."""
        if self.text:
            self.boundary_after[len(self.text) - 1] = True
        self.g = 0  # reset 'last' to root for a fresh example

    # ----- Serialization -----
    def to_state_dict(self) -> Dict:
        return {
            "b": self.b,
            "c": self.c,
            "d": self.d,
            "e": self.e,
            "g": self.g,
            "text_str": "".join(self.text),
            "boundary_after": self.boundary_after,
        }

    @classmethod
    def from_state_dict(cls, obj: Dict) -> "ROSACharPredictor":
        inst = cls()
        inst.b = obj["b"]
        inst.c = obj["c"]
        inst.d = obj["d"]
        inst.e = obj["e"]
        inst.g = obj.get("g", 0)
        inst.text = list(obj.get("text_str", ""))
        inst.boundary_after = obj.get("boundary_after", [False] * len(inst.text))
        if len(inst.boundary_after) < len(inst.text):
            inst.boundary_after += [False] * (len(inst.text) - len(inst.boundary_after))
        return inst


# ===============================
# 3) Walk a frozen SAM (inference)
# ===============================

def _advance_state(b, c, d, v: int, ch: str) -> int:
    """Walk the trained SAM with character `ch` (no learning). Returns new state v'."""
    while v != -1 and ch not in b[v]:
        v = c[v]
    if v == -1:
        return b[0].get(ch, 0)
    return b[v][ch]


def _predict_from_state(
    b, c, d, e, train_text: str, v: int, boundary_after: Optional[List[bool]] = None
) -> Optional[str]:
    """Deterministic ROSA next-char from current state v using rightmost indices.
    Refuses to cross an example boundary.
    """
    u = v
    n = len(train_text)
    while u != -1:
        i = e[u]
        j = i + 1
        if d[u] > 0 and 0 <= j < n:
            if boundary_after is not None and 0 <= i < len(boundary_after) and boundary_after[i]:
                u = c[u]
                continue
            return train_text[j]
        u = c[u]
    return None


# ============================================
# 4) Fallback LM that never crosses boundaries
# ============================================

class ROSAFallbackLM:
    """Character LM using Witten–Bell interpolation down the SAM suffix chain.
    IMPORTANT: counts are constructed per-example; no cross-boundary pairs.
    """
    def __init__(self, sam: ROSACharPredictor, examples: List[str], max_order: Optional[int] = None, show_progress: bool = True):
        self.b, self.c, self.d, self.e = sam.b, sam.c, sam.d, sam.e
        self.max_order = max_order
        joined = "".join(examples)
        self.alphabet = sorted(set(joined)) or ['\n']
        self.freq: List[Dict[str, int]] = [defaultdict(int) for _ in range(len(self.b))]
        self.unigram = Counter(joined) if joined else Counter({'\n': 1})
        self._cache: Dict[int, Dict[str, float]] = {}
        self._build_counts_examples_fast(examples, show_progress=show_progress)

    def _build_counts_examples_fast(self, examples: List[str], show_progress: bool = True) -> None:
        """Optimized: single pass per example, update-exclusion at the longest context,
        optionally clamped by `max_order`.
        """
        total_pairs = sum(max(0, len(seg) - 1) for seg in examples if seg)
        pbar = tqdm(total=total_pairs, desc="Training fallback LM (pairs)", disable=not show_progress, leave=False)

        b, c, d = self.b, self.c, self.d
        freq = self.freq
        max_order = self.max_order

        for seg in examples:
            if not seg:
                continue
            v = 0  # root
            for i in range(len(seg) - 1):
                ch = seg[i]
                # inline _advance_state
                u = v
                while u != -1 and ch not in b[u]:
                    u = c[u]
                if u == -1:
                    v = b[0].get(ch, 0)
                else:
                    v = b[u][ch]

                ctx = v
                if max_order is not None:
                    while ctx != -1 and d[ctx] > max_order:
                        ctx = c[ctx]
                    if ctx == -1:
                        ctx = 0
                nxt = seg[i + 1]
                freq[ctx][nxt] += 1
                pbar.update(1)

        pbar.close()

        self._propagate_counts_up_suffix_links()
        # invalidate any cached distributions built before propagation
        self._cache.clear()

    def ensure_capacity(self) -> None:
        """Make sure freq has one bucket per SAM state (called after SAM grows)."""
        missing = len(self.b) - len(self.freq)
        if missing > 0:
            self.freq.extend(defaultdict(int) for _ in range(missing))

    def observe_pair(self, ctx_state: int, next_ch: str, *, propagate: bool = True) -> None:
        """
        Online update: record one observation of (ctx_state -> next_ch).
        If propagate=True we mirror the offline build's suffix-up accumulation.
        """
        # Keep alphabet fixed during online learning (simple path).
        if next_ch not in self.alphabet:
            return  # or raise if you prefer a hard failure

        self.ensure_capacity()
        self.freq[ctx_state][next_ch] += 1

        if propagate:
            u = self.c[ctx_state]
            while u != -1:
                self.freq[u][next_ch] += 1
                u = self.c[u]

        # Invalidate memoized distributions for this chain
        u = ctx_state
        while u != -1:
            self._cache.pop(u, None)
            u = self.c[u]

    def _probs_for_state(self, v: int) -> Dict[str, float]:
        """Witten–Bell interpolation down suffix links, memoized by state."""
        if v in self._cache:
            return self._cache[v]

        # Build suffix chain (optionally truncated by max_order)
        chain = []
        u = v
        while u != -1:
            if self.max_order is not None and self.d[u] > self.max_order:
                u = self.c[u]
                continue
            chain.append(u)
            u = self.c[u]

        residual = 1.0
        probs: Dict[str, float] = {}

        def add_counts(state: int, scale: float):
            if scale <= 0.0:
                return
            total = sum(self.freq[state].values())
            if total == 0:
                return
            inv_total = 1.0 / total
            for ch, cnt in self.freq[state].items():
                probs[ch] = probs.get(ch, 0.0) + scale * (cnt * inv_total)

        for state in chain:
            N = sum(self.freq[state].values())
            T = len(self.freq[state])
            if N == 0:
                continue
            lam = N / (N + T) if T > 0 else 1.0  # Witten–Bell
            add_counts(state, residual * lam)
            residual *= (1.0 - lam)

        # Unigram fallback
        total_uni = sum(self.unigram.values())
        if total_uni > 0 and residual > 0.0:
            inv_total = 1.0 / total_uni
            for ch, cnt in self.unigram.items():
                probs[ch] = probs.get(ch, 0.0) + residual * (cnt * inv_total)

        s = sum(probs.values())
        if s > 0:
            inv_s = 1.0 / s
            for k in list(probs.keys()):
                probs[k] *= inv_s
        else:
            u = 1.0 / max(1, len(self.alphabet))
            probs = {ch: u for ch in self.alphabet}

        self._cache[v] = probs
        return probs
    
    def _propagate_counts_up_suffix_links(self) -> None:
        """
        After filling self.freq only at the longest contexts, push counts up the
        suffix-link tree so every shorter context has aggregated counts.
        Process states in decreasing d[v] so children flow into parents.
        """
        order = sorted(range(len(self.b)), key=lambda v: self.d[v], reverse=True)
        for v in order:
            p = self.c[v]
            if p < 0:      # root has no parent
                continue
            if not self.freq[v]:
                continue
            dv = self.freq[v]
            dp = self.freq[p]
            for ch, cnt in dv.items():
                dp[ch] += cnt

    @staticmethod
    def _sample_from_dist(
        dist: Dict[str, float],
        temperature: float = 1.0,
        top_p: Optional[float] = 0.9,
        top_k: Optional[int] = None,
    ) -> str:
        if temperature <= 0:
            temperature = 1e-6
        items = sorted(dist.items(), key=lambda x: x[1], reverse=True)
        if top_k is not None and top_k > 0:
            items = items[:max(1, top_k)]
        if top_p is not None:
            cum, cut = 0.0, []
            for ch, p in items:
                cum += p
                cut.append((ch, p))
                if cum >= top_p:
                    break
            items = cut or items[:1]
        logits = [math.log(max(p, 1e-12)) / temperature for _, p in items]
        m = max(logits)
        exps = [math.exp(z - m) for z in logits]
        Z = sum(exps)
        probs = [x / Z for x in exps]
        idx = random.choices(range(len(items)), weights=probs, k=1)[0]
        return items[idx][0]

    # ----- Serialization -----
    def to_state_dict(self) -> Dict:
        freq_plain = [{k: int(v) for k, v in d.items()} for d in tqdm(self.freq, leave=False)]
        return {
            "alphabet": self.alphabet,
            "unigram": {k: int(v) for k, v in self.unigram.items()},
            "freq": freq_plain,
            "max_order": self.max_order,
        }

    @classmethod
    def from_state_dict(cls, sam: ROSACharPredictor, obj: Dict) -> "ROSAFallbackLM":
        inst = cls.__new__(cls)  # bypass __init__
        inst.b, inst.c, inst.d, inst.e = sam.b, sam.c, sam.d, sam.e
        inst.max_order = obj.get("max_order", None)
        inst.alphabet = obj["alphabet"]
        inst.unigram = Counter({k: int(v) for k, v in obj["unigram"].items()})
        inst.freq = [defaultdict(int, {k: int(v) for k, v in d.items()}) for d in tqdm(obj["freq"], leave=False)]
        inst._cache = {}
        return inst


# ============================================
# 5) Mixed generation: deterministic + fallback
# ============================================

def _generate_mixed(
    sam: ROSACharPredictor,
    lm: ROSAFallbackLM,
    prompt: str,
    max_steps: int = 200,
    always_fallback = False,
    stop_at: Optional[str] = None,
    fallback_temperature: float = 1.0,
    fallback_top_p: Optional[float] = 0.9,
    fallback_top_k: Optional[int] = 50,
) -> str:
    b, c, d, e = sam.b, sam.c, sam.d, sam.e
    train_text = "".join(sam.text)
    v = 0
    for ch in prompt:
        v = _advance_state(b, c, d, v, ch)

    out: List[str] = []
    for _ in range(max_steps):
        # 1) Deterministic ROSA
        ch = None
        if not always_fallback:
            ch = _predict_from_state(b, c, d, e, train_text, v, boundary_after=sam.boundary_after)
        if ch is None:
            dist = lm._probs_for_state(v)
            ch = ROSAFallbackLM._sample_from_dist(
                dist, temperature=fallback_temperature, top_p=fallback_top_p, top_k=fallback_top_k
            )
        out.append(ch)
        if stop_at is not None and ch == stop_at:
            break
        v = _advance_state(b, c, d, v, ch)

    return "".join(out)


# ============================================
# 6) High-level wrapper class
# ============================================

MODEL_MAGIC = "rosa_pb_v2"  # v2 **only**

class ROSAPlus:
    """High-level model wrapper exposing a clean API.

    Typical usage:
        m = ROSAPlus(max_order=1048576, use_eot=True, eot_char="\u0004", seed=0)
        m.train_example("hello world\u0004")  # or skip EOT if you set use_eot=False
        m.build_lm()  # required before generation / probs
        out = m.generate("he", steps=20, temperature=0.7)
        p = m.next_char_prob("he", "l")
        m.save("model.bin")
        m2 = ROSAPlus.load("model.bin")
    """

    def __init__(self, *, max_order: Optional[int] = 1048576, use_eot: bool = True, eot_char: str = "\u0004", seed: int = 0):
        if not isinstance(eot_char, str) or len(eot_char) != 1:
            raise ValueError("eot_char must be a single character.")
        self.max_order = max_order
        self.use_eot = use_eot
        self.eot_char = eot_char
        self.seed = seed
        random.seed(seed)

        self.sam = ROSACharPredictor()
        self._examples: List[str] = []
        self.lm: Optional[ROSAFallbackLM] = None  # built later
        self.neural: Optional["ROSAGRUAdapter"] = None  # optional GRU adapter

    # ---- Training API ----
    def train_example(self, example: str) -> None:
        """Train on a single example string. Appends EOT if `use_eot` and it's not present as the final char."""
        if not example:
            return
        if self.use_eot:
            if example[-1] != self.eot_char:
                example = example + self.eot_char
        self._examples.append(example)
        for ch in tqdm(example):
            self.sam.feed(ch)
        self.sam.mark_boundary()

    def fit_from_examples(self, examples: List[str], *, show_progress: bool = True) -> None:
        """Convenience: train on many examples and build LM."""
        total_chars = sum(len(ex) for ex in examples)
        with tqdm(total=total_chars, desc="Training SAM (ROSA) over chars", disable=not show_progress) as pbar:
            for ex in examples:
                self.train_example(ex)  # this already appends EOT if enabled
                pbar.update(len(ex) if ex else 0)
        self.build_lm(show_progress=show_progress)

    def build_lm(self, *, show_progress: bool = True) -> None:
        if not self._examples:
            raise RuntimeError("No training examples available. Use train_example() first.")
        self.lm = ROSAFallbackLM(self.sam, self._examples, max_order=self.max_order, show_progress=show_progress)

    # ---- Optional neural adapter integration ----
    def attach_gru_adapter(self, adapter: "ROSAGRUAdapter") -> None:
        """Attach a trained (or untrained) GRU adapter."""
        self.neural = adapter

    def train_gru_adapter(
        self,
        *,
        examples,
        emb_dim: int = 128,
        hidden_dim: int = 256,
        num_layers: int = 1,
        dropout: float = 0.0,
        combine: str = "poe",
        beta: float = 1.0,
        device: Optional[str] = None,
        epochs: int = 1,
        lr: float = 1e-3,
        max_tokens_per_step: int = 4096*10,
        clip_grad: float = 1.0,
        show_progress: bool = True,
        # NEW:
        online_sam: bool = False,
        online_lm: bool = False,
        propagate_updates: bool = True,
    ) -> "ROSAGRUAdapter":
        if self.lm is None:
            raise RuntimeError("Fallback LM not built. Call build_lm() first.")
        adapter = ROSAGRUAdapter(
            self.lm.alphabet,
            emb_dim=emb_dim,
            hidden_dim=hidden_dim,
            num_layers=num_layers,
            dropout=dropout,
            combine=combine,
            beta=beta,
            device=device,
        )
        adapter.fit(
            self.sam,
            self.lm,
            examples,
            eot_char=self.eot_char,
            epochs=epochs,
            lr=lr,
            max_tokens_per_step=max_tokens_per_step,
            clip_grad=clip_grad,
            show_progress=show_progress,
            online_sam=online_sam,
            online_lm=online_lm,
            propagate_updates=propagate_updates,
        )
        self.neural = adapter
        return adapter

    # ---- Inference API ----
    def generate(
        self,
        prompt: str,
        *,
        steps: int = 200,
        always_fallback = False,
        stop_at: Optional[str] = None,
        temperature: float = 0.5,
        top_p: Optional[float] = 0.9,
        top_k: Optional[int] = 50,
        use_gru_adapter: bool = True,  # enable/disable neural reweighting
    ) -> str:
        """
        Generate a continuation from `prompt`.

        Order of operations per token:
        1) Try deterministic ROSA next-char (won't cross boundaries).
        2) Otherwise sample from fallback LM; if a GRU adapter is attached,
            refine the LM distribution first. The adapter consumes one base
            distribution per token; we *always* call step_with_char(ch) after
            emitting a char—this becomes a no-op if refine_distribution() was
            called in the same tick (the adapter guards against double steps).
        """
        if self.lm is None:
            raise RuntimeError("Fallback LM not built. Call build_lm() after training.")

        if stop_at is None and self.use_eot:
            stop_at = self.eot_char
        if top_k is not None and top_k <= 0:
            top_k = None

        # Unpack SAM internals
        b, c, d, e = self.sam.b, self.sam.c, self.sam.d, self.sam.e
        train_text = "".join(self.sam.text)

        # Walk SAM with the prompt to get the starting state
        v = 0
        for ch in prompt:
            v = _advance_state(b, c, d, v, ch)

        # Prime GRU with the prompt by consuming base distributions along the path
        if use_gru_adapter and self.neural is not None:
            try:
                self.neural.reset()
                v_prime = 0
                for ch in prompt:
                    v_prime = _advance_state(b, c, d, v_prime, ch)
                    base = self.lm._probs_for_state(v_prime)
                    if hasattr(self.neural, "step_with_dist"):
                        self.neural.step_with_dist(base)
            except Exception:
                # Priming is best-effort; fall back silently if adapter errors
                pass

        out: List[str] = []
        for _ in range(steps):
            # 1) Deterministic ROSA (if allowed)
            ch = None
            if not always_fallback:
                ch = _predict_from_state(
                    b, c, d, e, train_text, v, boundary_after=self.sam.boundary_after
                )

            # 2) Otherwise sample from (possibly GRU-refined) fallback LM
            if ch is None:
                base_dist = self.lm._probs_for_state(v)
                if use_gru_adapter and self.neural is not None and hasattr(self.neural, "refine_distribution"):
                    try:
                        dist = self.neural.refine_distribution(base_dist)  # consumes one step
                    except Exception:
                        dist = base_dist
                else:
                    dist = base_dist

                ch = ROSAFallbackLM._sample_from_dist(
                    dist, temperature=temperature, top_p=top_p, top_k=top_k
                )

            out.append(ch)
            if stop_at is not None and ch == stop_at:
                break

            # Advance SAM (and adapter) with the emitted char
            v = _advance_state(b, c, d, v, ch)
            if use_gru_adapter and self.neural is not None and hasattr(self.neural, "step_with_char"):
                try:
                    # If refine_distribution() already stepped this tick, this is a no-op.
                    self.neural.step_with_char(ch)
                except Exception:
                    pass

        return "".join(out)

    def get_dist(self, context: str, deterministic=False) -> Dict[str, float]:
        """Return a dict of next-char probabilities given `context`.
        If deterministic=True: Get direct prediction instead of probability (no fallback)
        """
        if self.lm is None:
            raise RuntimeError("Fallback LM not built. Call build_lm() after training.")
        b, c, d, e = self.sam.b, self.sam.c, self.sam.d, self.sam.e
        train_text = "".join(self.sam.text)
        v = 0
        for ch in context:
            v = _advance_state(b, c, d, v, ch)
            if deterministic:
                det = _predict_from_state(b, c, d, e, train_text, v, boundary_after=self.sam.boundary_after)
                if det is not None:
                    return {det: 1.0}
        base = self.lm._probs_for_state(v)
        # If neural adapter is attached, offer a reweighted distribution helper-style
        if self.neural is not None and hasattr(self.neural, "refine_distribution"):
            try:
                return self.neural.refine_distribution(base)
            except Exception:
                return base
        return base

    # ---- Persistence ----
    def save(self, path: str) -> None:
        if self.lm is None:
            raise RuntimeError("Fallback LM not built. Call build_lm() before saving.")
        payload = {
            "magic": MODEL_MAGIC,
            "sam": self.sam.to_state_dict(),
            "lm": self.lm.to_state_dict(),
            "meta": {
                "use_eot": self.use_eot,
                "eot_char": self.eot_char,
                "max_order": self.max_order,
                "seed": self.seed,
            },
            # Note: neural adapter is NOT serialized here by default to avoid
            # adding a framework dependency into the model blob.
        }
        with open(path, "wb") as f:
            f.write(_dumps(payload))

    @classmethod
    def load(cls, path: str) -> "ROSAPlus":
        with open(path, "rb") as f:
            payload = _loads(f.read())
        magic = payload.get("magic")
        if magic != MODEL_MAGIC:
            raise ValueError(f"Unrecognized or unsupported model magic in {path}: {magic} (v2 only)")
        meta = payload.get("meta", {})
        inst = cls(
            max_order=meta.get("max_order", 1048576),
            use_eot=meta.get("use_eot", True),
            eot_char=meta.get("eot_char", "\u0004"),
            seed=meta.get("seed", 0),
        )
        inst.sam = ROSACharPredictor.from_state_dict(payload["sam"])  # type: ignore[arg-type]
        inst.lm = ROSAFallbackLM.from_state_dict(inst.sam, payload["lm"])  # type: ignore[arg-type]
        return inst

    # ---- Utilities ----
    @staticmethod
    def decode_escape(s: str) -> str:
        try:
            return s.encode("utf-8").decode("unicode_escape")
        except Exception:
            return s