File size: 19,164 Bytes
503dfd6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Record encoding, block-causal mask and row form. Ported from Kev's kev/model.py (Apache-2.0, see NOTICE) and
extended with pause tokens, option echo, rationale (tier-2) branches and plain-text rows
(docs/reasoning-architecture.md, section 1).

Decision record -> one packed sequence; every (question, tier) branch has its own seg and restarts after the state:

    <state> state tokens | <q> instr <opt> o1 </opt> ... <opt> oK </opt>  SUFFIX  | <q> ... | ...

    SUFFIX, tier 0/1           <pause>*N <decide>                                   readout: </opt> of o1..oK
    SUFFIX, tier 0/1 + echo    <pause>*N <opt> k1 </opt> ... <opt> kK </opt> <decide>      echoed </opt>
    SUFFIX, tier 2             <think> r1..rm <decide>                              readout: </opt> of o1..oK
    SUFFIX, tier 2 + echo      <think> r1..rm <opt> k1 </opt> ... <opt> kK </opt> <decide> echoed </opt>

k_j is option j's short key (`echo_key`). Two echo modes (the doc's table, section 1):
  siblings    attention-only bases: the spans share position ids (each starts right after the last pause / rationale
              token), each sees everything before the echo block plus itself (`sib`, enforced by the packed mask and by
              the row and fork forms' 4D masks), <decide> sits after the longest span and sees all of them, so the block
              is order-invariant
  sequential  hybrids (recurrent layers cannot be masked): consecutive positions, the spans see each other in option
              order; the same layout as Kev's echo without option isolation, which fine-tunes the Qwen3.5 checkpoints
A question with a rationale emits its tier-0/1 branch always and a tier-2 branch with probability p_tier2.

seg[i]      : 0 for state tokens, k for the tokens of branch k (k >= 1)
pos[i]      : position ids; every branch restarts right after the state, so branches are interchangeable
sib[i]      : j + 1 on the tokens of echo span j in siblings mode, else 0
decide_idx  : index of each branch's <decide>; opt_idx: index of each option's scored </opt> (echoed when echo is on)
lm_labels   : next-token targets (the model shifts them): text rows everywhere; decision rows only on the rationale
              (r1..rm, and <decide> as the stop target after rm, placed on the first echo token in echo mode); a
              corrupted rationale (or a self-sampled one marked correct: false) gets none. Never on state, instructions,
              options, pauses or echo.
qidx        : index in rec["questions"] of each branch's question
tiers, rat_corrupt, rat_targets : per branch 0/1/2 and "rationale without LM targets"; LM targets in the encoding

Text record -> one causal row with seg all 0 and lm_labels = ids (the model shifts them).
"""
import hashlib
import re
import torch

from .tokens import Delimiters

OPT_NONE, OPT_DECIDE, OPT_PAUSE, OPT_THINK, OPT_ECHO = -1, -2, -3, -4, -5
IGNORE = -100


class ContextOverflow(ValueError):
    """A record does not fit the context (state, branch or packed limit)."""


_SPECIAL_RE = re.compile(r"<\|([A-Za-z0-9_]+)\|>")


def user_tokens(tok, delims: Delimiters, text):
    """Tokenize caller text so it can never produce a delimiter or pause token (option boundaries are unforgeable).
    `<|name|>` is rewritten to `<¦name¦>` (Kev) and any literal delimiter/pause string gets the same treatment."""
    text = _SPECIAL_RE.sub(r"<¦\1¦>", text)
    for name in delims.names.values():
        if name in text:
            text = text.replace(name, "<¦" + name[1:-1] + "¦>")
    return tok(text, add_special_tokens=False).input_ids


def render(v, indent=0):
    """Flatten a JSON state (str | object | array) into the text the model sees, exactly as Kev's serving path does
    (kev.api.render), so a mid-trained base sees the same state text the fine-tune and the server will give it.
    trainer/data wraps ~30 % of hf states as {"document": ...}, {"ticket": ...} or a chat list."""
    pad = "  " * indent
    if v is None: return ""
    if isinstance(v, (str, int, float, bool)): return str(v)
    if isinstance(v, list): return "\n".join(f"{pad}- {render(x, indent + 1).lstrip()}" for x in v)
    return "\n".join(f"{pad}{k}:\n{render(x, indent + 1)}" if isinstance(x, (dict, list)) else f"{pad}{k}: {render(x)}" for k, x in v.items())


_SLUG = re.compile(r"[a-z0-9][a-z0-9_.-]{0,39}")


def option_texts(q):
    """The option strings exactly as the fine-tune and the server render them (t2m_kev.convert.to_kev_question ->
    kev.api.to_record): choice options that are all distinct short slugs stay as they are, anything else becomes
    `a: text`, `b: text`, ... (`o27: ...` past 26); noul is `no` / `yes` (`no: text` when the options are not literally
    no/yes); score levels are the level texts. Mid-training on the same surface form is what lets the head and the
    delimiter embeddings carry over into the fine-tune.
    A question may carry `option_texts`, the strings already rendered (eval.predictors.MidtrainPredictor passes
    kev.api.to_record's, so Kev-format keys such as `c: ...` or `none_of_these: ...` reach the model unchanged)."""
    if q.get("option_texts") is not None:
        return [str(o) for o in q["option_texts"]]
    opts = [str(o) for o in q["options"]]
    if q["type"] == "choice":
        if len(set(opts)) == len(opts) and all(_SLUG.fullmatch(o) for o in opts):
            return opts
        keys = [chr(ord("a") + i) if len(opts) <= 26 else f"o{i + 1}" for i in range(len(opts))]
        return [f"{k}: {o}" if o else k for k, o in zip(keys, opts)]
    if q["type"] == "noul" and opts != ["no", "yes"]:
        return [f"no: {opts[0]}" if opts[0] else "no", f"yes: {opts[1]}" if opts[1] else "yes"]
    return opts


def echo_key(q, j, text):
    """The short key option j is echoed by (reasoning-architecture.md, token layouts): the slug or letter `option_texts`
    puts before the colon, `no`/`yes` for noul, the level index for score."""
    if q["type"] == "score":
        return str(j)
    if q["type"] == "noul":
        return ("no", "yes")[j]
    return text.split(": ", 1)[0]       # as kev.model.echo_keys falls back to


def rationale_of(q):
    """(text, lm) of a question's rationale, or None. Accepts the trainer-JSONL string form (`rationale` plus an optional
    `rationale_corrupt: true`) and the object form of reasoning-architecture.md section 3 ({"text", "corrupt", "correct",
    ...}). lm is False for a corrupted rationale and for a self-sampled one marked `correct: false`: those train the head
    only. An object carrying `ids` (a decoded rationale, see rationale_ids) counts even when empty (an immediate stop)."""
    r = q.get("rationale")
    if isinstance(r, dict):
        text, corrupt, correct = r.get("text"), r.get("corrupt", q.get("rationale_corrupt", False)), r.get("correct", True)
        if isinstance(r.get("ids"), list):   # decoded token ids (predict, tier 2): used as they are, possibly none
            return (text if isinstance(text, str) else ""), not (bool(corrupt) or correct is False)
    elif isinstance(r, str):
        text, corrupt, correct = r, q.get("rationale_corrupt", False), True
    else:
        return None
    if not isinstance(text, str) or not text.strip():
        return None
    return text, not (bool(corrupt) or correct is False)


def rationale_ids(tok, delims, q, text):
    """The rationale's token ids: `rationale.ids` when the object form carries them (a rationale decoded by the model at
    tier 2, midtrain.predict: re-tokenizing its text need not give back the decoded ids), else user_tokens(text)."""
    r = q.get("rationale")
    if isinstance(r, dict) and isinstance(r.get("ids"), list):
        return [int(t) for t in r["ids"]]
    return user_tokens(tok, delims, text)


def _draw(key, p):
    """A deterministic Bernoulli(p) draw from a string key (stable across processes, unlike hash())."""
    if p >= 1:
        return True
    if p <= 0:
        return False
    return int.from_bytes(hashlib.blake2b(key.encode(), digest_size=8).digest(), "big") / 2 ** 64 < p


def echo_mode(hybrid):
    """The echo layout a base supports: sibling spans need a maskable (attention-only) backbone."""
    return "sequential" if hybrid else "siblings"


def decision_kwargs(cfg, hybrid=False):
    """encode_decision keyword arguments for the config's echo / rationale switches; {} when both are off (the pilot
    configs), so their encodings are unchanged."""
    kw = {}
    if (cfg.get("pause") or {}).get("echo_options"):
        kw["echo"] = echo_mode(hybrid)
    r = cfg.get("rationale") or {}
    if r.get("enabled"):
        kw["rationale"] = {"p_tier2": float(r.get("p_tier2", 0.3)), "max_tokens": int(r.get("max_tokens", 512)),
                           "seed": str((cfg.get("train") or {}).get("seed", 0))}
    return kw


def encode_decision(tok, delims: Delimiters, rec, n_pause=0, max_state=1024, max_branch=1024, max_packed=4096, strict=False,
                    echo=False, rationale=None, rng_key=""):
    """Encode one decision record. Returns a list of packed encodings: usually one; several when the branches do not fit
    one packed sequence (each chunk repeats the state). Raises ContextOverflow when a tier-0/1 branch does not fit; a
    tier-2 branch that does not fit (or whose rationale exceeds rationale["max_tokens"]) is dropped and counted.

    echo: False, or repeat each option's key after the pauses / rationale and score the echoed </opt> (option echo,
      arm C2) in mode "siblings" (True) or "sequential" (see the module docstring; `echo_mode(model.hybrid)` picks).
    rationale: None (questions' rationales are ignored) or {"p_tier2", "max_tokens", "seed"}: a question with a
      rationale also emits a tier-2 branch with probability p_tier2, drawn deterministically from (seed, rng_key, the
      question); needs the `think` delimiter role."""
    state_tokens = user_tokens(tok, delims, render(rec["state"]))
    if strict and len(state_tokens) + 1 > max_state:
        raise ContextOverflow(f"state exceeds {max_state} tokens: {len(state_tokens) + 1}")
    S = [delims["state"]] + state_tokens[: max_state - 1]
    q_id, o_id, c_id, d_id, p_id = (delims[r] for r in ("q", "opt", "opt_end", "decide", "pause"))
    if echo not in (False, None, True, "siblings", "sequential"):
        raise ValueError(f"echo must be False, True, 'siblings' or 'sequential', not {echo!r}")
    sibling = echo in (True, "siblings")
    if rationale is not None and "think" not in delims.ids:
        raise ValueError("rationale training needs the `think` delimiter (install_delimiters(tok, {'think': '<think>'}))")
    branches, dropped = [], 0      # (ids, opt, pos offsets, sib, labels, readout offsets, question, its index, tier, corrupt)
    for qi, q in enumerate(rec["questions"]):
        instr = [q_id] + user_tokens(tok, delims, q["instructions"])
        texts = option_texts(q)
        spans = [[o_id] + user_tokens(tok, delims, o) + [c_id] for o in texts]
        prefix = instr + [t for sp in spans for t in sp]
        prefix_opt = [OPT_NONE] * len(instr) + [j for j, sp in enumerate(spans) for _ in sp]
        ends, cursor = [], len(instr)
        for sp in spans:
            cursor += len(sp); ends.append(cursor - 1)
        echo_spans = [[o_id] + user_tokens(tok, delims, echo_key(q, j, t)) + [c_id] for j, t in enumerate(texts)] if echo else None
        suffixes = [([p_id] * n_pause, [OPT_PAUSE] * n_pause, [IGNORE] * n_pause, 1 if n_pause else 0, False, False)]
        rat = rationale_of(q) if rationale is not None else None
        if rat is not None and _draw(f"{rationale.get('seed', '')}:{rng_key}:{qi}:{q.get('id')}:{q['instructions']}:{rat[0]}", rationale["p_tier2"]):
            r_ids = rationale_ids(tok, delims, q, rat[0])
            if len(r_ids) > int(rationale.get("max_tokens", 512)):
                dropped += 1
            else:
                mid = [delims["think"]] + r_ids
                suffixes.append((mid, [OPT_THINK] * len(mid), [IGNORE] + (r_ids if rat[1] else [IGNORE] * len(r_ids)), 2, rat[1], not rat[1]))
        for mid, mid_opt, mid_lab, tier, lm, corrupt in suffixes:
            br, bopt, blab = prefix + mid, prefix_opt + mid_opt, [IGNORE] * len(prefix) + mid_lab
            boff, bsib = list(range(len(br))), [0] * len(br)
            if echo:
                e0, readout = len(br), []
                for j, sp in enumerate(echo_spans):
                    start = e0 if sibling else len(br)
                    boff += range(start, start + len(sp)); bsib += [j + 1 if sibling else 0] * len(sp)
                    br += sp; bopt += [OPT_ECHO] * len(sp); blab += [IGNORE] * len(sp)
                    readout.append(len(br) - 1)
                d_off = e0 + max(len(sp) for sp in echo_spans) if sibling else len(br)
                if lm:
                    blab[e0] = d_id    # the stop is learned, the echo is not: h[rm] predicts <decide> though <opt> follows
            else:
                readout, d_off = list(ends), len(br)
            br.append(d_id); bopt.append(OPT_DECIDE); boff.append(d_off); bsib.append(0)
            blab.append(d_id if lm and not echo else IGNORE)
            if len(br) > max_branch or len(S) + len(br) > max_packed:
                if tier == 2:
                    dropped += 1
                    continue
                if len(br) > max_branch:
                    raise ContextOverflow(f"branch too long: {len(br)} tokens (limit {max_branch})")
                raise ContextOverflow(f"state + branch exceeds {max_packed} packed tokens")
            branches.append((br, bopt, boff, bsib, blab, readout, q, qi, tier, corrupt))
    out, i = [], 0
    while i < len(branches):
        ids, seg, pos, opt, sib, lab = list(S), [0] * len(S), list(range(len(S))), [OPT_NONE] * len(S), [0] * len(S), [IGNORE] * len(S)
        decide_idx, opt_idx, targets, qtypes, qids, qidx, tiers, corrupt = [], [], [], [], [], [], [], []
        k = 0
        while i < len(branches) and len(ids) + len(branches[i][0]) <= max_packed:
            br, bopt, boff, bsib, blab, readout, q, qi, tier, bad = branches[i]; k += 1
            base, p0 = len(ids), len(S)
            ids += br; seg += [k] * len(br); pos += [p0 + o for o in boff]; opt += bopt; sib += bsib; lab += blab
            decide_idx.append(base + len(br) - 1); opt_idx.append([base + e for e in readout])
            targets.append(q["target"]); qtypes.append(q["type"]); qids.append(q.get("id", f"q{qi + 1}")); qidx.append(qi)
            tiers.append(tier); corrupt.append(bad)
            i += 1
        out.append({"kind": "decision", "ids": ids, "seg": seg, "pos": pos, "opt": opt, "sib": sib, "decide_idx": decide_idx, "opt_idx": opt_idx,
                    "targets": targets, "qtypes": qtypes, "qids": qids, "qidx": qidx, "tiers": tiers, "rat_corrupt": corrupt, "lm_labels": lab,
                    "rat_targets": sum(l != IGNORE for l in lab), "n_pause": n_pause, "echo": ("siblings" if sibling else "sequential") if echo else False,
                    "tier2_dropped": dropped if not out else 0, "state_truncated": len(state_tokens) + 1 > max_state})
    return out


def encode_text(tok, delims: Delimiters, text, max_len=2048):
    """One causal row for the plain LM loss. Documents end with EOS so the model learns boundaries; BOS is added when the
    tokenizer has one (Llama-style bases)."""
    ids = user_tokens(tok, delims, text)
    if tok.bos_token_id is not None and getattr(tok, "add_bos_token", False):
        ids = [tok.bos_token_id] + ids
    ids = ids[: max_len - 1]
    if tok.eos_token_id is not None:
        ids = ids + [tok.eos_token_id]
    L = len(ids)
    return {"kind": "text", "ids": ids, "seg": [0] * L, "pos": list(range(L)), "opt": [OPT_NONE] * L, "sib": [0] * L, "decide_idx": [], "opt_idx": [],
            "targets": [], "qtypes": [], "qids": [], "qidx": [], "tiers": [], "rat_corrupt": [], "lm_labels": list(ids), "rat_targets": 0, "n_pause": 0,
            "echo": False, "tier2_dropped": 0, "state_truncated": False}


def branch_mask_batch(segs, device, dtype=torch.float32, length=None, sibs=None):
    """Batched block-causal mask, additive [B,1,L,L], right-padded to the longest sequence (Kev).
    attend(i,j) iff j<=i and (seg[j]==0 or seg[j]==seg[i]); pads (-1) are masked keys; the diagonal is always kept
    so no row is fully masked. A row whose seg is all zero gets the plain causal mask (text rows).
    sibs (option echo): tokens of two different echo spans (sib > 0, unequal) never see each other."""
    L = max(max(len(s) for s in segs), length or 0)
    s = torch.full((len(segs), L), -1, device=device)
    for b, seg in enumerate(segs):
        s[b, : len(seg)] = torch.tensor(seg, device=device)
    causal = torch.tril(torch.ones(L, L, dtype=torch.bool, device=device))
    same = (s[:, None, :] == s[:, :, None]) | (s[:, None, :] == 0)
    valid_key = (s != -1)[:, None, :]
    allow = causal[None] & same & valid_key
    if sibs is not None:
        g = torch.zeros((len(segs), L), dtype=torch.long, device=device)
        for b, sb in enumerate(sibs):
            g[b, : len(sb)] = torch.tensor(sb, device=device)
        allow = allow & ~((g[:, :, None] > 0) & (g[:, None, :] > 0) & (g[:, :, None] != g[:, None, :]))
    allow = allow | torch.eye(L, dtype=torch.bool, device=device)[None]
    return torch.zeros(len(segs), L, L, dtype=dtype, device=device).masked_fill(~allow, torch.finfo(dtype).min)[:, None]


def rows_of(enc):
    """Split a packed decision encoding into (state_ids, state_pos, rows); rows[k] = {"ids", "pos", "labels", "sib",
    "decide", "opts"} holds branch k's tokens (state-continuing positions), its LM labels, echo siblings and readout
    offsets within the branch. state + rows[k] as one causal row equals the packed block-causal form for that branch
    on any architecture (Kev); with echo siblings, on attention-only bases, given the sibling mask."""
    seg = enc["seg"]; Ls = seg.count(0)
    labels, sib = enc["lm_labels"], enc.get("sib") or [0] * len(seg)
    rows, start = [], Ls
    for k, (d, oi) in enumerate(zip(enc["decide_idx"], enc["opt_idx"]), start=1):
        end = d + 1
        if seg[start] != k or seg[end - 1] != k:
            raise ValueError("branch layout mismatch")
        rows.append({"ids": enc["ids"][start:end], "pos": enc["pos"][start:end], "labels": labels[start:end], "sib": sib[start:end],
                     "decide": d - start, "opts": [o - start for o in oi]})
        start = end
    return enc["ids"][:Ls], enc["pos"][:Ls], rows


def as_rows(enc):
    """The row form of any encoding: a text row is itself; a decision encoding becomes one causal row per branch,
    each = state + branch, with absolute readout indices. Returns list of (ids, pos, lm_labels, decide, opts, sib)."""
    if enc["kind"] == "text":
        return [(enc["ids"], enc["pos"], enc["lm_labels"], None, None, enc.get("sib") or [0] * len(enc["ids"]))]
    S, Sp, rows = rows_of(enc)
    return [(S + r["ids"], Sp + r["pos"], [IGNORE] * len(S) + r["labels"], len(S) + r["decide"], [len(S) + o for o in r["opts"]], [0] * len(S) + r["sib"])
            for r in rows]