File size: 20,397 Bytes
095660c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Stable prompt and artifact contract shared by the Studio and product SFT."""

from __future__ import annotations

import hashlib
import json
import re
from pathlib import Path
from typing import Any, Mapping, Sequence


CHAT_PROMPT_FORMAT = "smolgpt-fables-chat-v3"
SMOLLM3_CHAT_PROMPT_FORMAT = "smolgpt-fables-smollm3-chat-v5"
RAW_PROMPT_FORMAT = "smolgpt-fables-raw-v1"
PRODUCT_CONTEXT_LENGTH = 2048
SMOLLM3_PRODUCT_CONTEXT_LENGTH = 4096
CHAT_SYSTEM_PROMPT = (
    "You are SmolGPT-Fables. Output only the requested finished Markdown story "
    "continuation. Begin exactly with `### Scene 01:`. Emit exactly the requested "
    "number of consecutive, zero-padded `### Scene NN:` sections. Those scene "
    "headings are the only headings allowed in the output: never emit an H1, H2, "
    "any other H3, an H4 or deeper heading, or any other section before, between, "
    "or after them. Follow the canvas exactly and copy every required name and "
    "detail verbatim. Stop immediately after the requested final scene. Never "
    "repeat or quote the story title, metadata, canvas, `## Story`, or any instruction."
)
CHAT_PROMPT_TRANSFORM_VERSION = "scene-output-contract-v1"
CHAT_STORY_BOUNDARY = "## Story\n\n"
CHAT_OUTPUT_CONTRACT_TEMPLATE = (
    "Output contract (follow exactly): use only these H3 heading prefixes, in this "
    "order: {headings}. Write exactly {scene_count} scenes; emit no other heading "
    "or section; stop immediately after completing `### Scene {final_scene:02d}:`.\n\n"
)
SMOLLM3_PROMPT_TRANSFORM_VERSION = "natural-fable-scene-contract-v3"
SMOLLM3_STORY_BOUNDARY = "## Story\n\n"
SMOLLM3_SCENE_WORD_RANGE = (45, 115)
SMOLLM3_SYSTEM_PROMPT = (
    "You are SmolGPT-Fables. Write a vivid, complete fable from the user's canvas. "
    "Output only the finished story continuation. Begin with `### Scene 01:` and "
    "emit exactly the requested consecutive, zero-padded scene sections. A scene "
    "heading may include a short title after the colon. Use no other Markdown "
    "heading. Copy every required name, setting, and unusual detail verbatim. Make "
    "each character's described role, personality, and desire affect what they do. "
    "Write concrete action and dialogue instead of summarizing instructions. Keep "
    "each scene concise, make every scene change the situation, and resolve the "
    "ending target inside the final scene. Stop immediately after the final sentence; "
    "never add notes, analysis, a moral label, an ending section, or quoted canvas "
    "text. /no_think"
)
_SCENE_COUNT_PATTERN = re.compile(r"(?m)^- Scene Count: ([0-9]+)$")
_TARGET_SCENES_PATTERN = re.compile(r"(?m)^- Target scenes: ([0-9]+)$")
_MIN_SCENE_COUNT = 1
_MAX_SCENE_COUNT = 6
SUPPORTED_PROMPT_FORMATS = frozenset(
    {CHAT_PROMPT_FORMAT, SMOLLM3_CHAT_PROMPT_FORMAT, RAW_PROMPT_FORMAT}
)
CHAT_PROMPT_FORMATS = frozenset(
    {CHAT_PROMPT_FORMAT, SMOLLM3_CHAT_PROMPT_FORMAT}
)
COMMON_ARTIFACT_FILES = ("config.json", "tokenizer.json")
CUSTOM_CODE_FILES = ("configuration_smolgpt.py", "modeling_smolgpt.py")


def _render_output_contract(scene_count: int) -> str:
    headings = ", ".join(
        f"`### Scene {index:02d}:`" for index in range(1, scene_count + 1)
    )
    return CHAT_OUTPUT_CONTRACT_TEMPLATE.format(
        headings=headings,
        scene_count=scene_count,
        final_scene=scene_count,
    )


def prompt_contract_sha256() -> str:
    payload = {
        "format": CHAT_PROMPT_FORMAT,
        "system": CHAT_SYSTEM_PROMPT,
        "messages": ["system", "user", "assistant"],
        "assistant_only_loss": True,
        "prompt_transform": {
            "version": CHAT_PROMPT_TRANSFORM_VERSION,
            "scene_count_pattern": _SCENE_COUNT_PATTERN.pattern,
            "target_scenes_pattern": _TARGET_SCENES_PATTERN.pattern,
            "scene_count_range": [_MIN_SCENE_COUNT, _MAX_SCENE_COUNT],
            "story_boundary": CHAT_STORY_BOUNDARY,
            "rendered_output_contracts": {
                str(scene_count): _render_output_contract(scene_count)
                for scene_count in range(_MIN_SCENE_COUNT, _MAX_SCENE_COUNT + 1)
            },
        },
    }
    encoded = json.dumps(payload, sort_keys=True, separators=(",", ":")).encode()
    return hashlib.sha256(encoded).hexdigest()


def smollm3_render_output_contract(scene_count: int) -> str:
    """Render the exact SmolLM3 v5 output contract used for SFT."""

    if not _MIN_SCENE_COUNT <= scene_count <= _MAX_SCENE_COUNT:
        raise ValueError("v4 scene count must be between 1 and 6")
    headings = ", ".join(
        f"`### Scene {index:02d}:`"
        for index in range(1, scene_count + 1)
    )
    minimum, maximum = SMOLLM3_SCENE_WORD_RANGE
    return (
        "Output contract (follow exactly):\n"
        f"- Use these scene prefixes in order: {headings}.\n"
        f"- Write exactly {scene_count} scenes and {minimum}-{maximum} words per scene.\n"
        "- Use no heading except those scene headings.\n"
        "- Copy every item on `Must include` verbatim into the story.\n"
        "- Show the character-role details through decisions, action, or dialogue.\n"
        f"- Resolve the ending target in Scene {scene_count:02d} and stop.\n\n"
    )


def _canonical_smollm3_scene_count(prompt: str) -> int:
    scene_counts = _SCENE_COUNT_PATTERN.findall(prompt)
    target_counts = _TARGET_SCENES_PATTERN.findall(prompt)
    if len(scene_counts) != 1 or len(target_counts) != 1:
        raise ValueError("v4 prompt needs one Scene Count and one Target scenes line")
    scene_count = int(scene_counts[0])
    target_count = int(target_counts[0])
    if not _MIN_SCENE_COUNT <= scene_count <= _MAX_SCENE_COUNT:
        raise ValueError("v4 scene count must be between 1 and 6")
    if scene_count != target_count:
        raise ValueError("v4 Scene Count and Target scenes must match")
    return scene_count


def smollm3_transform_prompt(prompt: str) -> str:
    """Apply the exact SmolLM3 v5 prompt transform used for SFT."""

    scene_count = _canonical_smollm3_scene_count(prompt)
    if not prompt.endswith(SMOLLM3_STORY_BOUNDARY):
        raise ValueError("v4 prompt must end at the canonical Story boundary")
    return (
        prompt[: -len(SMOLLM3_STORY_BOUNDARY)]
        + smollm3_render_output_contract(scene_count)
        + SMOLLM3_STORY_BOUNDARY
    )


def smollm3_chat_messages(
    prompt: str,
    completion: str | None = None,
) -> list[dict[str, str]]:
    messages = [
        {"role": "system", "content": SMOLLM3_SYSTEM_PROMPT},
        {"role": "user", "content": smollm3_transform_prompt(prompt)},
    ]
    if completion is not None:
        messages.append({"role": "assistant", "content": completion})
    return messages


def smollm3_prompt_contract_sha256() -> str:
    """Hash the exact SmolLM3 v5 prompt contract used for SFT."""

    payload = {
        "format": SMOLLM3_CHAT_PROMPT_FORMAT,
        "system": SMOLLM3_SYSTEM_PROMPT,
        "messages": ["system", "user", "assistant"],
        "assistant_only_loss": True,
        "thinking": False,
        "context_length": SMOLLM3_PRODUCT_CONTEXT_LENGTH,
        "transform_version": SMOLLM3_PROMPT_TRANSFORM_VERSION,
        "story_boundary": SMOLLM3_STORY_BOUNDARY,
        "scene_count_pattern": _SCENE_COUNT_PATTERN.pattern,
        "target_scenes_pattern": _TARGET_SCENES_PATTERN.pattern,
        "scene_word_range": list(SMOLLM3_SCENE_WORD_RANGE),
        "rendered_contracts": {
            str(count): smollm3_render_output_contract(count)
            for count in range(_MIN_SCENE_COUNT, _MAX_SCENE_COUNT + 1)
        },
    }
    encoded = json.dumps(payload, sort_keys=True, separators=(",", ":")).encode()
    return hashlib.sha256(encoded).hexdigest()


def raw_prompt_contract_sha256() -> str:
    payload = {
        "format": RAW_PROMPT_FORMAT,
        "messages": ["raw-markdown-prompt"],
        "bos_prefix": True,
    }
    encoded = json.dumps(payload, sort_keys=True, separators=(",", ":")).encode()
    return hashlib.sha256(encoded).hexdigest()


def product_context_length_for_prompt_format(prompt_format: str) -> int | None:
    if prompt_format == CHAT_PROMPT_FORMAT:
        return PRODUCT_CONTEXT_LENGTH
    if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT:
        return SMOLLM3_PRODUCT_CONTEXT_LENGTH
    if prompt_format == RAW_PROMPT_FORMAT:
        return None
    raise ValueError(f"unsupported prompt format: {prompt_format}")


def _has_model_weights(root: Path) -> bool:
    return (root / "model.safetensors").is_file() or (
        root / "model.safetensors.index.json"
    ).is_file()


def validate_transformers_artifact(root: Path) -> tuple[Mapping[str, Any], str]:
    """Validate either the legacy custom export or a standard Transformers LM."""

    missing = [name for name in COMMON_ARTIFACT_FILES if not (root / name).is_file()]
    if missing:
        raise ValueError("model repository is missing: " + ", ".join(missing))
    if not _has_model_weights(root):
        raise ValueError("model repository is missing Safetensors weights")
    try:
        config = json.loads((root / "config.json").read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError) as exc:
        raise ValueError(f"could not read config.json: {exc}") from exc
    if not isinstance(config, Mapping):
        raise ValueError("config.json must contain a JSON object")
    auto_map = config.get("auto_map")
    if isinstance(auto_map, Mapping) and auto_map.get("AutoModelForCausalLM"):
        missing_code = [name for name in CUSTOM_CODE_FILES if not (root / name).is_file()]
        if missing_code:
            raise ValueError("custom model repository is missing: " + ", ".join(missing_code))
        return config, "custom"
    architectures = config.get("architectures")
    if not isinstance(architectures, list) or not all(
        isinstance(value, str) and value for value in architectures
    ):
        raise ValueError("standard model config needs a non-empty architectures list")
    if not isinstance(config.get("model_type"), str) or not config["model_type"]:
        raise ValueError("standard model config needs model_type")
    return config, "standard"


def _expected_prompt_contract_sha256(prompt_format: str) -> str:
    if prompt_format == CHAT_PROMPT_FORMAT:
        return prompt_contract_sha256()
    if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT:
        return smollm3_prompt_contract_sha256()
    if prompt_format == RAW_PROMPT_FORMAT:
        return raw_prompt_contract_sha256()
    raise ValueError(f"unsupported prompt format: {prompt_format}")


def _manifest_prompt_binding(
    manifest: Mapping[str, Any],
    *,
    artifact_kind: str,
) -> tuple[str, str] | None:
    """Read legacy top-level or SmolLM3 nested bindings without ambiguity."""

    top_format_present = "prompt_format" in manifest
    top_hash_present = "prompt_contract_sha256" in manifest
    nested_present = "prompt_contract" in manifest
    if not top_format_present and not top_hash_present and not nested_present:
        if artifact_kind == "custom":
            return None
        raise ValueError("training manifest has no supported prompt_format")
    if top_format_present != top_hash_present:
        raise ValueError("training manifest prompt binding is incomplete")

    bindings: list[tuple[str, str]] = []
    if top_format_present:
        bindings.append(
            (manifest.get("prompt_format"), manifest.get("prompt_contract_sha256"))
        )
    if nested_present:
        nested = manifest.get("prompt_contract")
        if (
            not isinstance(nested, Mapping)
            or set(nested) != {"format", "sha256", "thinking"}
            or nested.get("thinking") is not False
        ):
            raise ValueError("training manifest nested prompt contract is invalid")
        bindings.append((nested.get("format"), nested.get("sha256")))
    if any(
        not isinstance(value, str)
        or not value
        or not isinstance(digest, str)
        or re.fullmatch(r"[0-9a-f]{64}", digest) is None
        for value, digest in bindings
    ):
        raise ValueError("training manifest prompt binding is invalid")
    if any(binding != bindings[0] for binding in bindings[1:]):
        raise ValueError("training manifest prompt bindings conflict")
    return bindings[0]


def prompt_format_for_artifact(root: Path, artifact_kind: str, tokenizer: Any) -> str:
    manifest_path = root / "training_manifest.json"
    if manifest_path.is_file():
        manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
        if not isinstance(manifest, Mapping):
            raise ValueError("training manifest must contain a JSON object")
        binding = _manifest_prompt_binding(manifest, artifact_kind=artifact_kind)
        if binding is None:
            # Checked-in pre-contract SmolGPT exports are unambiguously the raw
            # Markdown/BOS path. Any partially declared contract still fails.
            return RAW_PROMPT_FORMAT
        value, contract_hash = binding
        if value not in SUPPORTED_PROMPT_FORMATS:
            raise ValueError("training manifest has no supported prompt_format")
        if artifact_kind == "standard" and value not in CHAT_PROMPT_FORMATS:
            raise ValueError("standard product manifest must use the chat prompt format")
        if artifact_kind == "custom" and value == SMOLLM3_CHAT_PROMPT_FORMAT:
            raise ValueError("SmolLM3 chat-v5 requires a standard model artifact")
        expected_sha = _expected_prompt_contract_sha256(value)
        if contract_hash != expected_sha:
            raise ValueError("training manifest prompt contract hash does not match runtime")
        if value in CHAT_PROMPT_FORMATS and not getattr(
            tokenizer, "chat_template", None
        ):
            raise ValueError("chat prompt format requires a tokenizer chat_template")
        return str(value)
    if artifact_kind == "standard" and getattr(tokenizer, "chat_template", None):
        return CHAT_PROMPT_FORMAT
    return RAW_PROMPT_FORMAT


def _canonical_scene_count(prompt: str) -> int:
    scene_counts = _SCENE_COUNT_PATTERN.findall(prompt)
    target_counts = _TARGET_SCENES_PATTERN.findall(prompt)
    if len(scene_counts) != 1:
        raise ValueError(
            "chat-v3 prompt must contain exactly one canonical '- Scene Count: N' line"
        )
    if len(target_counts) != 1:
        raise ValueError(
            "chat-v3 prompt must contain exactly one canonical '- Target scenes: N' line"
        )
    scene_count = int(scene_counts[0])
    target_count = int(target_counts[0])
    if not _MIN_SCENE_COUNT <= scene_count <= _MAX_SCENE_COUNT:
        raise ValueError("chat-v3 Scene Count must be between 1 and 6")
    if not _MIN_SCENE_COUNT <= target_count <= _MAX_SCENE_COUNT:
        raise ValueError("chat-v3 Target scenes must be between 1 and 6")
    if scene_count != target_count:
        raise ValueError("chat-v3 Scene Count and Target scenes must match")
    return scene_count


def _transform_chat_prompt(prompt: str) -> str:
    scene_count = _canonical_scene_count(prompt)
    if not prompt.endswith(CHAT_STORY_BOUNDARY):
        raise ValueError("chat-v3 prompt must end at the canonical '## Story' boundary")
    return (
        prompt[: -len(CHAT_STORY_BOUNDARY)]
        + _render_output_contract(scene_count)
        + CHAT_STORY_BOUNDARY
    )


def chat_messages(prompt: str, completion: str | None = None) -> list[dict[str, str]]:
    messages = [
        {"role": "system", "content": CHAT_SYSTEM_PROMPT},
        {"role": "user", "content": _transform_chat_prompt(prompt)},
    ]
    if completion is not None:
        messages.append({"role": "assistant", "content": completion})
    return messages


def _chat_messages_for_format(
    prompt: str,
    completion: str | None,
    prompt_format: str,
) -> list[dict[str, str]]:
    if prompt_format == CHAT_PROMPT_FORMAT:
        return chat_messages(prompt, completion)
    if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT:
        return smollm3_chat_messages(prompt, completion)
    raise ValueError(f"unsupported chat prompt format: {prompt_format}")


def _apply_runtime_chat_template(
    tokenizer: Any,
    messages: list[dict[str, str]],
    *,
    prompt_format: str,
    add_generation_prompt: bool,
) -> Any:
    kwargs: dict[str, Any] = {
        "add_generation_prompt": add_generation_prompt,
        "tokenize": True,
    }
    if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT:
        kwargs["enable_thinking"] = False
    return tokenizer.apply_chat_template(messages, **kwargs)


def _flatten_token_ids(values: Any, *, context: str) -> list[int]:
    """Normalize chat-template outputs across supported Transformers versions."""

    if isinstance(values, Mapping):
        if "input_ids" not in values:
            raise ValueError(f"{context} returned no input_ids")
        values = values["input_ids"]
    if hasattr(values, "tolist"):
        values = values.tolist()
    if not isinstance(values, Sequence) or isinstance(values, (str, bytes, bytearray)):
        raise ValueError(f"{context} returned unsupported token IDs")
    normalized = list(values)
    if normalized and isinstance(normalized[0], Sequence) and not isinstance(
        normalized[0], (str, bytes, bytearray)
    ):
        if len(normalized) != 1:
            raise ValueError(f"{context} returned more than one token sequence")
        normalized = list(normalized[0])
    try:
        return [int(value) for value in normalized]
    except (TypeError, ValueError) as exc:
        raise ValueError(f"{context} returned non-integer token IDs") from exc


def generation_prompt_ids(tokenizer: Any, prompt: str, prompt_format: str) -> list[int]:
    if prompt_format in CHAT_PROMPT_FORMATS:
        values = _apply_runtime_chat_template(
            tokenizer,
            _chat_messages_for_format(prompt, None, prompt_format),
            prompt_format=prompt_format,
            add_generation_prompt=True,
        )
        context = (
            "SmolLM3 v5 generation template"
            if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT
            else "chat generation template"
        )
        return _flatten_token_ids(values, context=context)
    if prompt_format != RAW_PROMPT_FORMAT:
        raise ValueError(f"unsupported prompt format: {prompt_format}")
    return [
        int(tokenizer.bos_token_id),
        *(
            int(value)
            for value in tokenizer.encode(prompt, add_special_tokens=False)
        ),
    ]


def assistant_training_ids(
    tokenizer: Any,
    prompt: str,
    completion: str,
    *,
    max_length: int,
    prompt_format: str = CHAT_PROMPT_FORMAT,
) -> tuple[list[int], list[int]]:
    """Create one chat sequence with loss masked through the assistant header."""

    if prompt_format not in CHAT_PROMPT_FORMATS:
        raise ValueError("assistant training requires a supported chat prompt format")
    prefix = generation_prompt_ids(tokenizer, prompt, prompt_format)
    full = _flatten_token_ids(
        _apply_runtime_chat_template(
            tokenizer,
            _chat_messages_for_format(prompt, completion, prompt_format),
            prompt_format=prompt_format,
            add_generation_prompt=False,
        ),
        context=(
            "SmolLM3 v5 training template"
            if prompt_format == SMOLLM3_CHAT_PROMPT_FORMAT
            else "chat training template"
        ),
    )
    if full[: len(prefix)] != prefix:
        raise ValueError("chat template assistant prefix is not stable")
    if len(full) > max_length:
        raise ValueError(
            f"chat-formatted example has {len(full)} tokens; maximum is {max_length}"
        )
    if len(full) <= len(prefix):
        raise ValueError("chat-formatted example has no assistant completion tokens")
    labels = [-100] * len(prefix) + full[len(prefix) :]
    return full, labels


def aggregate_sha256(paths: Sequence[Path], root: Path) -> str:
    records = []
    for path in sorted(paths):
        records.append(
            {
                "path": path.relative_to(root).as_posix(),
                "sha256": hashlib.sha256(path.read_bytes()).hexdigest(),
            }
        )
    payload = json.dumps(records, sort_keys=True, separators=(",", ":")).encode()
    return hashlib.sha256(payload).hexdigest()