| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """On-demand trace builder for parser engine testing and benchmarks. |
| 4 | |
| 5 | Generates token sequences programmatically from model-agnostic scenario |
| 6 | definitions. Each model format handler knows how to render scenarios |
| 7 | into the model's output format, tokenize them with correct special token |
| 8 | IDs, and compute expected parse outputs. |
| 9 | |
| 10 | Every generated sample is self-validated by replaying it through the |
| 11 | real parser before being returned. |
| 12 | """ |
| 13 | |
| 14 | from __future__ import annotations |
| 15 | |
| 16 | import functools |
| 17 | import json |
| 18 | from collections.abc import Callable |
| 19 | from dataclasses import dataclass, replace |
| 20 | from typing import Any |
| 21 | |
| 22 | from tests.parser.engine.replay_harness import ( |
| 23 | MockTokenizer, |
| 24 | Sample, |
| 25 | assert_parse_output, |
| 26 | collect_output, |
| 27 | replay_streaming, |
| 28 | ) |
| 29 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 30 | ChatCompletionToolsParam, |
| 31 | ) |
| 32 | from vllm.parser import plamo3 |
| 33 | from vllm.parser.engine.registered_adapters import ( |
| 34 | DeepSeekV4Parser, |
| 35 | DeepSeekV32Parser, |
| 36 | DeepSeekV41Parser, |
| 37 | Gemma4Parser, |
| 38 | Glm47MoeParser, |
| 39 | GraniteParser, |
| 40 | GraniteThinkingParser, |
| 41 | InklingParser, |
| 42 | KimiK2Parser, |
| 43 | MiMoParser, |
| 44 | MinimaxM2Parser, |
| 45 | NemotronV3Parser, |
| 46 | Plamo3Parser, |
| 47 | Qwen3Parser, |
| 48 | SeedOssParser, |
| 49 | Step3p5Parser, |
| 50 | ) |
| 51 | |
| 52 | # ── Data structures ────────────────────────────────────────────────── |
| 53 | |
| 54 | |
| 55 | @dataclass |
| 56 | class ToolCallSpec: |
| 57 | name: str |
| 58 | arguments: dict[str, Any] |
| 59 | |
| 60 | |
| 61 | @dataclass |
| 62 | class Scenario: |
| 63 | id: str |
| 64 | description: str |
| 65 | reasoning: str | None = None |
| 66 | content: str | None = None |
| 67 | tool_calls: list[ToolCallSpec] | None = None |
| 68 | after_tool_response: bool = False |
| 69 | |
| 70 | |
| 71 | # ── Scenarios ──────────────────────────────────────────────────────── |
| 72 | |
| 73 | _READ_TOOL = ToolCallSpec("read_file", {"path": "/tmp/test.txt"}) |
| 74 | _BASH_TOOL = ToolCallSpec( |
| 75 | "bash", {"command": "hostname", "description": "Get hostname"} |
| 76 | ) |
| 77 | _WEATHER_TOOL = ToolCallSpec( |
| 78 | "get_weather", |
| 79 | {"city": "Dallas", "state": "TX", "unit": "fahrenheit"}, |
| 80 | ) |
| 81 | _COMPLEX_TOOL = ToolCallSpec( |
| 82 | "search", |
| 83 | { |
| 84 | "query": "vllm parser", |
| 85 | "filters": {"language": "python", "min_stars": 100}, |
| 86 | "tags": ["ml", "inference"], |
| 87 | "limit": 10, |
| 88 | "verbose": True, |
| 89 | }, |
| 90 | ) |
| 91 | |
| 92 | SCENARIOS: list[Scenario] = [ |
| 93 | Scenario( |
| 94 | id="think-then-tool", |
| 95 | description="Reasoning then single tool call", |
| 96 | reasoning="Let me check the file.", |
| 97 | tool_calls=[_READ_TOOL], |
| 98 | ), |
| 99 | Scenario( |
| 100 | id="think-then-parallel-tools", |
| 101 | description="Reasoning then two parallel tool calls", |
| 102 | reasoning="I need to run both commands.", |
| 103 | tool_calls=[_BASH_TOOL, _WEATHER_TOOL], |
| 104 | ), |
| 105 | Scenario( |
| 106 | id="think-then-content", |
| 107 | description="Reasoning then content response", |
| 108 | reasoning="Let me think about this carefully.", |
| 109 | content="The answer is 42.", |
| 110 | ), |
| 111 | Scenario( |
| 112 | id="content-only", |
| 113 | description="Plain content response without reasoning", |
| 114 | content="Hello! How can I help you today?", |
| 115 | ), |
| 116 | Scenario( |
| 117 | id="tool-only", |
| 118 | description="Tool call without reasoning", |
| 119 | tool_calls=[_READ_TOOL], |
| 120 | ), |
| 121 | Scenario( |
| 122 | id="complex-json-args", |
| 123 | description="Tool call with nested objects, arrays, numbers, booleans", |
| 124 | reasoning="This needs a complex query.", |
| 125 | tool_calls=[_COMPLEX_TOOL], |
| 126 | ), |
| 127 | Scenario( |
| 128 | id="whitespace-before-tool", |
| 129 | description="Whitespace-only content before tool call", |
| 130 | content="\n\n", |
| 131 | tool_calls=[_WEATHER_TOOL], |
| 132 | ), |
| 133 | Scenario( |
| 134 | id="think-content-tool", |
| 135 | description="Reasoning, content, then tool call", |
| 136 | reasoning="Let me analyze and then fetch data.", |
| 137 | content="Checking the weather now.", |
| 138 | tool_calls=[_WEATHER_TOOL], |
| 139 | ), |
| 140 | Scenario( |
| 141 | id="think-whitespace-tool", |
| 142 | description="Reasoning, whitespace-only gap, then tool call", |
| 143 | reasoning="Let me check the file contents.", |
| 144 | content="\n\n", |
| 145 | tool_calls=[_READ_TOOL], |
| 146 | ), |
| 147 | Scenario( |
| 148 | id="empty-reasoning-content", |
| 149 | description="Empty reasoning section followed by content", |
| 150 | reasoning="", |
| 151 | content="The epoch timestamp is 1779111346.", |
| 152 | ), |
| 153 | Scenario( |
| 154 | id="tool-after-tool-response", |
| 155 | description="Tool call immediately after tool response (agentic flow)", |
| 156 | tool_calls=[_READ_TOOL], |
| 157 | after_tool_response=True, |
| 158 | ), |
| 159 | Scenario( |
| 160 | id="empty-tool-block", |
| 161 | description="Empty tool block followed by content (edge case recovery)", |
| 162 | content="Content after empty tools.", |
| 163 | tool_calls=[], |
| 164 | ), |
| 165 | ] |
| 166 | |
| 167 | |
| 168 | # ── Tokenization ───────────────────────────────────────────────────── |
| 169 | |
| 170 | |
| 171 | def _word_split(text: str) -> list[str]: |
| 172 | """Split text into word-like tokens, preserving all characters.""" |
| 173 | if not text: |
| 174 | return [] |
| 175 | parts: list[str] = [] |
| 176 | current = "" |
| 177 | for ch in text: |
| 178 | if ch in " \t\n\r" and current and current[-1] not in " \t\n\r": |
| 179 | parts.append(current) |
| 180 | current = ch |
| 181 | else: |
| 182 | current += ch |
| 183 | if current: |
| 184 | parts.append(current) |
| 185 | return parts |
| 186 | |
| 187 | |
| 188 | def _tokenize( |
| 189 | segments: list[tuple[str, bool]], |
| 190 | vocab: dict[str, int], |
| 191 | start_id: int = 100, |
| 192 | ) -> list[tuple[int, str]]: |
| 193 | """Build token list from segments. |
| 194 | |
| 195 | Each segment is ``(text, is_special)``. Special segments use vocab |
| 196 | IDs; content segments are word-split with sequential IDs. |
| 197 | """ |
| 198 | tokens: list[tuple[int, str]] = [] |
| 199 | next_id = start_id |
| 200 | |
| 201 | for text, is_special in segments: |
| 202 | if not text: |
| 203 | continue |
| 204 | if is_special: |
| 205 | tid = vocab.get(text) |
| 206 | if tid is None: |
| 207 | raise ValueError(f"Special token {text!r} not in vocab") |
| 208 | tokens.append((tid, text)) |
| 209 | else: |
| 210 | for word in _word_split(text): |
| 211 | tokens.append((next_id, word)) |
| 212 | next_id += 1 |
| 213 | |
| 214 | return tokens |
| 215 | |
| 216 | |
| 217 | # ── Tool definitions ───────────────────────────────────────────────── |
| 218 | |
| 219 | |
| 220 | def _infer_schema(value: object) -> dict: |
| 221 | """Infer a JSON Schema from a Python value, recursing into dicts/lists.""" |
| 222 | if isinstance(value, bool): |
| 223 | return {"type": "boolean"} |
| 224 | if isinstance(value, int): |
| 225 | return {"type": "integer"} |
| 226 | if isinstance(value, float): |
| 227 | return {"type": "number"} |
| 228 | if isinstance(value, str): |
| 229 | return {"type": "string"} |
| 230 | if isinstance(value, dict): |
| 231 | return { |
| 232 | "type": "object", |
| 233 | "properties": {k: _infer_schema(v) for k, v in value.items()}, |
| 234 | } |
| 235 | if isinstance(value, list) and value: |
| 236 | return {"type": "array", "items": _infer_schema(value[0])} |
| 237 | if isinstance(value, list): |
| 238 | return {"type": "array"} |
| 239 | return {} |
| 240 | |
| 241 | |
| 242 | def _tool_defs(tool_calls: list[ToolCallSpec]) -> list[dict]: |
| 243 | """Generate OpenAI-style tool definitions from tool call specs.""" |
| 244 | seen: set[str] = set() |
| 245 | tools: list[dict] = [] |
| 246 | for tc in tool_calls: |
| 247 | if tc.name in seen: |
| 248 | continue |
| 249 | seen.add(tc.name) |
| 250 | properties = {k: _infer_schema(v) for k, v in tc.arguments.items()} |
| 251 | tools.append( |
| 252 | { |
| 253 | "type": "function", |
| 254 | "function": { |
| 255 | "name": tc.name, |
| 256 | "parameters": { |
| 257 | "type": "object", |
| 258 | "properties": properties, |
| 259 | }, |
| 260 | }, |
| 261 | } |
| 262 | ) |
| 263 | return tools |
| 264 | |
| 265 | |
| 266 | # ── Format handlers ────────────────────────────────────────────────── |
| 267 | |
| 268 | |
| 269 | def _expected_tc(scenario: Scenario) -> list[dict] | None: |
| 270 | if not scenario.tool_calls: |
| 271 | return None |
| 272 | return [{"name": tc.name, "arguments": tc.arguments} for tc in scenario.tool_calls] |
| 273 | |
| 274 | |
| 275 | def _expected_tools(scenario: Scenario) -> list[dict] | None: |
| 276 | return _tool_defs(scenario.tool_calls) if scenario.tool_calls else None |
| 277 | |
| 278 | |
| 279 | def _validate_sample(sample: Sample, parser_cls: type, **kwargs) -> None: |
| 280 | """Replay sample through the real parser and assert correctness.""" |
| 281 | tokenizer = MockTokenizer(vocab=dict(sample.vocab), tokens=sample.tokens) |
| 282 | parser = parser_cls(tokenizer, sample.tools, **kwargs) |
| 283 | deltas = replay_streaming( |
| 284 | parser, |
| 285 | sample.tokens, |
| 286 | chunk_size=1, |
| 287 | tools=sample.tools, |
| 288 | prompt_token_ids=sample.prompt_token_ids, |
| 289 | ) |
| 290 | output = collect_output(deltas) |
| 291 | assert_parse_output(output, sample) |
| 292 | |
| 293 | |
| 294 | def _validate_tools( |
| 295 | tools: list[dict] | None, |
| 296 | ) -> list[ChatCompletionToolsParam] | None: |
| 297 | if not tools: |
| 298 | return None |
| 299 | return [ChatCompletionToolsParam.model_validate(t) for t in tools] |
| 300 | |
| 301 | |
| 302 | def _make_sample( |
| 303 | sample_id: str, |
| 304 | description: str, |
| 305 | vocab: dict[str, int], |
| 306 | segments: list[tuple[str, bool]], |
| 307 | expected_reasoning: str | None, |
| 308 | expected_content: str | None, |
| 309 | expected_tool_calls: list[dict] | None, |
| 310 | tools: list[dict] | None, |
| 311 | chat_template_kwargs: dict | None = None, |
| 312 | prompt_token_ids: list[int] | None = None, |
| 313 | ) -> Sample: |
| 314 | tokens = _tokenize(segments, vocab) |
| 315 | return Sample( |
| 316 | id=sample_id, |
| 317 | description=description, |
| 318 | source="trace-builder", |
| 319 | vocab=dict(vocab), |
| 320 | tokens=tokens, |
| 321 | expected_reasoning=expected_reasoning, |
| 322 | expected_content=expected_content, |
| 323 | expected_tool_calls=expected_tool_calls, |
| 324 | tools=_validate_tools(tools), |
| 325 | chat_template_kwargs=chat_template_kwargs, |
| 326 | prompt_token_ids=prompt_token_ids, |
| 327 | ) |
| 328 | |
| 329 | |
| 330 | # ── Qwen3 (XML tool format, starts in REASONING) ──────────────────── |
| 331 | |
| 332 | _QWEN3_VOCAB: dict[str, int] = { |
| 333 | "<think>": 50, |
| 334 | "</think>": 51, |
| 335 | "<tool_call>": 60, |
| 336 | "</tool_call>": 61, |
| 337 | } |
| 338 | |
| 339 | |
| 340 | def _qwen3_arg_value(value: Any) -> str: |
| 341 | if isinstance(value, bool): |
| 342 | return "true" if value else "false" |
| 343 | if isinstance(value, (int, float)): |
| 344 | return str(value) |
| 345 | if isinstance(value, str): |
| 346 | return value |
| 347 | return json.dumps(value, ensure_ascii=False) |
| 348 | |
| 349 | |
| 350 | def _qwen3_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 351 | parts = [f"\n<function={tc.name}>"] |
| 352 | for key, value in tc.arguments.items(): |
| 353 | parts.append(f"\n<parameter={key}>{_qwen3_arg_value(value)}</parameter>") |
| 354 | parts.append("\n</function>\n") |
| 355 | return [ |
| 356 | ("<tool_call>", True), |
| 357 | ("".join(parts), False), |
| 358 | ("</tool_call>", True), |
| 359 | ] |
| 360 | |
| 361 | |
| 362 | def _qwen3_segments( |
| 363 | scenario: Scenario, |
| 364 | tool_segments: Callable[[ToolCallSpec], list[tuple[str, bool]]] = ( |
| 365 | _qwen3_tool_segments |
| 366 | ), |
| 367 | ) -> list[tuple[str, bool]]: |
| 368 | segs: list[tuple[str, bool]] = [] |
| 369 | if scenario.reasoning is not None: |
| 370 | segs.append((scenario.reasoning, False)) |
| 371 | if scenario.content is not None or scenario.tool_calls is not None: |
| 372 | segs.append(("</think>", True)) |
| 373 | if scenario.tool_calls is not None and not scenario.tool_calls: |
| 374 | segs.append(("<tool_call>", True)) |
| 375 | segs.append(("</tool_call>", True)) |
| 376 | if scenario.content is not None: |
| 377 | segs.append((scenario.content, False)) |
| 378 | if scenario.tool_calls: |
| 379 | for tc in scenario.tool_calls: |
| 380 | segs.extend(tool_segments(tc)) |
| 381 | return segs |
| 382 | |
| 383 | |
| 384 | def _qwen3_expected_content(scenario: Scenario) -> str | None: |
| 385 | if ( |
| 386 | scenario.content is not None |
| 387 | and scenario.tool_calls |
| 388 | and not scenario.content.strip() |
| 389 | ): |
| 390 | return "" |
| 391 | return scenario.content |
| 392 | |
| 393 | |
| 394 | def _build_qwen3( |
| 395 | scenario: Scenario, |
| 396 | name: str = "qwen3", |
| 397 | parser_cls: type = Qwen3Parser, |
| 398 | strip_trailing_ws: bool = False, |
| 399 | validate: bool = True, |
| 400 | tool_segments: Callable[[ToolCallSpec], list[tuple[str, bool]]] = ( |
| 401 | _qwen3_tool_segments |
| 402 | ), |
| 403 | ) -> Sample: |
| 404 | expected_reasoning: str | None |
| 405 | if scenario.reasoning is not None: |
| 406 | r = scenario.reasoning |
| 407 | if strip_trailing_ws: |
| 408 | r = r.rstrip() |
| 409 | expected_reasoning = r |
| 410 | else: |
| 411 | expected_reasoning = "" |
| 412 | |
| 413 | sample = _make_sample( |
| 414 | sample_id=f"{name}-{scenario.id}", |
| 415 | description=scenario.description, |
| 416 | vocab=_QWEN3_VOCAB, |
| 417 | segments=_qwen3_segments(scenario, tool_segments), |
| 418 | expected_reasoning=expected_reasoning, |
| 419 | expected_content=_qwen3_expected_content(scenario), |
| 420 | expected_tool_calls=_expected_tc(scenario), |
| 421 | tools=_expected_tools(scenario), |
| 422 | ) |
| 423 | if validate: |
| 424 | _validate_sample(sample, parser_cls) |
| 425 | return sample |
| 426 | |
| 427 | |
| 428 | # ── MiMo (compact Qwen3 XML, no newlines between tags) ─────────────── |
| 429 | |
| 430 | |
| 431 | def _mimo_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 432 | parts = [f"<function={tc.name}>"] |
| 433 | for key, value in tc.arguments.items(): |
| 434 | parts.append(f"<parameter={key}>{_qwen3_arg_value(value)}</parameter>") |
| 435 | parts.append("</function>") |
| 436 | return [ |
| 437 | ("<tool_call>", True), |
| 438 | ("".join(parts), False), |
| 439 | ("</tool_call>", True), |
| 440 | ] |
| 441 | |
| 442 | |
| 443 | def _build_mimo(scenario: Scenario, validate: bool = True) -> Sample: |
| 444 | return _build_qwen3( |
| 445 | scenario, |
| 446 | name="mimo", |
| 447 | parser_cls=MiMoParser, |
| 448 | validate=validate, |
| 449 | tool_segments=_mimo_tool_segments, |
| 450 | ) |
| 451 | |
| 452 | |
| 453 | # ── Step-3.5 (Qwen3 XML, trailing reasoning whitespace stripped) ──── |
| 454 | |
| 455 | |
| 456 | def _build_step3p5(scenario: Scenario, validate: bool = True) -> Sample: |
| 457 | return _build_qwen3( |
| 458 | scenario, |
| 459 | name="step3p5", |
| 460 | parser_cls=Step3p5Parser, |
| 461 | strip_trailing_ws=True, |
| 462 | validate=validate, |
| 463 | ) |
| 464 | |
| 465 | |
| 466 | # ── MiniMax M2 (XML invoke format, starts in REASONING) ────────────── |
| 467 | |
| 468 | _MINIMAX_M2_VOCAB: dict[str, int] = { |
| 469 | "<think>": 50, |
| 470 | "</think>": 51, |
| 471 | "<minimax:tool_call>": 60, |
| 472 | "</minimax:tool_call>": 61, |
| 473 | } |
| 474 | |
| 475 | |
| 476 | def _minimax_m2_arg_value(value: Any) -> str: |
| 477 | if isinstance(value, bool): |
| 478 | return "true" if value else "false" |
| 479 | if isinstance(value, (int, float)): |
| 480 | return str(value) |
| 481 | if isinstance(value, str): |
| 482 | return value |
| 483 | return json.dumps(value, ensure_ascii=False) |
| 484 | |
| 485 | |
| 486 | def _minimax_m2_tool_segments(tool_calls: list[ToolCallSpec]) -> list[tuple[str, bool]]: |
| 487 | segs: list[tuple[str, bool]] = [("<minimax:tool_call>", True)] |
| 488 | for tc in tool_calls: |
| 489 | segs.append((f'<invoke name="{tc.name}">', False)) |
| 490 | for key, value in tc.arguments.items(): |
| 491 | segs.append( |
| 492 | ( |
| 493 | f'<parameter name="{key}">' |
| 494 | f"{_minimax_m2_arg_value(value)}" |
| 495 | "</parameter>", |
| 496 | False, |
| 497 | ) |
| 498 | ) |
| 499 | segs.append(("</invoke>", False)) |
| 500 | segs.append(("</minimax:tool_call>", True)) |
| 501 | return segs |
| 502 | |
| 503 | |
| 504 | def _minimax_m2_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 505 | segs: list[tuple[str, bool]] = [] |
| 506 | if scenario.reasoning is not None: |
| 507 | segs.append((scenario.reasoning, False)) |
| 508 | if scenario.content is not None or scenario.tool_calls is not None: |
| 509 | segs.append(("</think>", True)) |
| 510 | if scenario.tool_calls is not None and not scenario.tool_calls: |
| 511 | segs.append(("<minimax:tool_call>", True)) |
| 512 | segs.append(("</minimax:tool_call>", True)) |
| 513 | if scenario.content is not None: |
| 514 | segs.append((scenario.content, False)) |
| 515 | if scenario.tool_calls: |
| 516 | segs.extend(_minimax_m2_tool_segments(scenario.tool_calls)) |
| 517 | return segs |
| 518 | |
| 519 | |
| 520 | def _build_minimax_m2(scenario: Scenario, validate: bool = True) -> Sample: |
| 521 | expected_reasoning: str | None |
| 522 | if scenario.reasoning is not None: |
| 523 | expected_reasoning = scenario.reasoning.rstrip() |
| 524 | else: |
| 525 | expected_reasoning = "" |
| 526 | |
| 527 | sample = _make_sample( |
| 528 | sample_id=f"minimax_m2-{scenario.id}", |
| 529 | description=scenario.description, |
| 530 | vocab=_MINIMAX_M2_VOCAB, |
| 531 | segments=_minimax_m2_segments(scenario), |
| 532 | expected_reasoning=expected_reasoning, |
| 533 | expected_content=_qwen3_expected_content(scenario), |
| 534 | expected_tool_calls=_expected_tc(scenario), |
| 535 | tools=_expected_tools(scenario), |
| 536 | ) |
| 537 | if validate: |
| 538 | _validate_sample(sample, MinimaxM2Parser) |
| 539 | return sample |
| 540 | |
| 541 | |
| 542 | # ── Gemma4 (channel reasoning, custom arg format) ──────────────────── |
| 543 | |
| 544 | _GEMMA4_VOCAB: dict[str, int] = { |
| 545 | "<|channel>": 50, |
| 546 | "<channel|>": 51, |
| 547 | "<|tool_call>": 48, |
| 548 | "<tool_call|>": 49, |
| 549 | '<|"|>': 52, |
| 550 | "<|turn>": 53, |
| 551 | "<|tool_response>": 54, |
| 552 | } |
| 553 | _GEMMA4_THOUGHT_PREFIX = "thought\n" |
| 554 | _GEMMA4_QUOTE = '<|"|>' |
| 555 | |
| 556 | |
| 557 | def _gemma4_value_segments(value: Any) -> list[tuple[str, bool]]: |
| 558 | """Render a value in Gemma4 arg format as segments.""" |
| 559 | if isinstance(value, str): |
| 560 | return [(_GEMMA4_QUOTE, True), (value, False), (_GEMMA4_QUOTE, True)] |
| 561 | if isinstance(value, bool): |
| 562 | return [("true" if value else "false", False)] |
| 563 | if isinstance(value, (int, float)): |
| 564 | return [(str(value), False)] |
| 565 | if isinstance(value, dict): |
| 566 | segs: list[tuple[str, bool]] = [("{", False)] |
| 567 | for i, (k, v) in enumerate(value.items()): |
| 568 | if i > 0: |
| 569 | segs.append((",", False)) |
| 570 | segs.append((f"{k}:", False)) |
| 571 | segs.extend(_gemma4_value_segments(v)) |
| 572 | segs.append(("}", False)) |
| 573 | return segs |
| 574 | if isinstance(value, list): |
| 575 | segs = [("[", False)] |
| 576 | for i, item in enumerate(value): |
| 577 | if i > 0: |
| 578 | segs.append((",", False)) |
| 579 | segs.extend(_gemma4_value_segments(item)) |
| 580 | segs.append(("]", False)) |
| 581 | return segs |
| 582 | return [(json.dumps(value, ensure_ascii=False), False)] |
| 583 | |
| 584 | |
| 585 | def _gemma4_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 586 | segs: list[tuple[str, bool]] = [ |
| 587 | ("<|tool_call>", True), |
| 588 | (f"call:{tc.name}", False), |
| 589 | ("{", False), |
| 590 | ] |
| 591 | for i, (key, value) in enumerate(tc.arguments.items()): |
| 592 | if i > 0: |
| 593 | segs.append((",", False)) |
| 594 | segs.append((f"{key}:", False)) |
| 595 | segs.extend(_gemma4_value_segments(value)) |
| 596 | segs.append(("}", False)) |
| 597 | segs.append(("<tool_call|>", True)) |
| 598 | return segs |
| 599 | |
| 600 | |
| 601 | def _gemma4_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 602 | segs: list[tuple[str, bool]] = [] |
| 603 | if scenario.reasoning is not None: |
| 604 | segs.append(("<|channel>", True)) |
| 605 | segs.append((_GEMMA4_THOUGHT_PREFIX, False)) |
| 606 | segs.append((scenario.reasoning, False)) |
| 607 | segs.append(("<channel|>", True)) |
| 608 | if scenario.tool_calls is not None and not scenario.tool_calls: |
| 609 | segs.append(("<|tool_call>", True)) |
| 610 | segs.append(("<tool_call|>", True)) |
| 611 | if scenario.content is not None: |
| 612 | segs.append((scenario.content, False)) |
| 613 | if scenario.tool_calls: |
| 614 | for tc in scenario.tool_calls: |
| 615 | segs.extend(_gemma4_tool_segments(tc)) |
| 616 | return segs |
| 617 | |
| 618 | |
| 619 | def _build_gemma4(scenario: Scenario, validate: bool = True) -> Sample: |
| 620 | prompt_token_ids = None |
| 621 | if scenario.after_tool_response: |
| 622 | prompt_token_ids = [_GEMMA4_VOCAB["<|tool_response>"]] |
| 623 | sample = _make_sample( |
| 624 | sample_id=f"gemma4-{scenario.id}", |
| 625 | description=scenario.description, |
| 626 | vocab=_GEMMA4_VOCAB, |
| 627 | segments=_gemma4_segments(scenario), |
| 628 | expected_reasoning=scenario.reasoning, |
| 629 | expected_content=_qwen3_expected_content(scenario), |
| 630 | expected_tool_calls=_expected_tc(scenario), |
| 631 | tools=_expected_tools(scenario), |
| 632 | prompt_token_ids=prompt_token_ids, |
| 633 | ) |
| 634 | if validate: |
| 635 | _validate_sample(sample, Gemma4Parser) |
| 636 | return sample |
| 637 | |
| 638 | |
| 639 | def _build_nemotron_v3(scenario: Scenario, validate: bool = True) -> Sample: |
| 640 | return _build_qwen3( |
| 641 | scenario, |
| 642 | name="nemotron_v3", |
| 643 | parser_cls=NemotronV3Parser, |
| 644 | strip_trailing_ws=True, |
| 645 | validate=validate, |
| 646 | ) |
| 647 | |
| 648 | |
| 649 | def _build_granite_thinking(scenario: Scenario, validate: bool = True) -> Sample: |
| 650 | """Granite 4.2: the Nemotron V3 grammar plus leading-newline stripping on |
| 651 | content after ``</think>``.""" |
| 652 | sample = _build_qwen3( |
| 653 | scenario, |
| 654 | name="granite_thinking_parser", |
| 655 | parser_cls=GraniteThinkingParser, |
| 656 | strip_trailing_ws=True, |
| 657 | validate=False, |
| 658 | ) |
| 659 | sample = replace( |
| 660 | sample, |
| 661 | expected_content=( |
| 662 | sample.expected_content.lstrip("\n") |
| 663 | if sample.expected_content |
| 664 | else sample.expected_content |
| 665 | ), |
| 666 | content_lstrip="\n", |
| 667 | ) |
| 668 | if validate: |
| 669 | _validate_sample(sample, GraniteThinkingParser) |
| 670 | return sample |
| 671 | |
| 672 | |
| 673 | # ── Seed-OSS (Qwen3 XML grammar with Seed wrapper tokens) ──────────── |
| 674 | |
| 675 | _SEED_OSS_VOCAB: dict[str, int] = { |
| 676 | "<seed:think>": 50, |
| 677 | "</seed:think>": 51, |
| 678 | "<seed:tool_call>": 60, |
| 679 | "</seed:tool_call>": 61, |
| 680 | } |
| 681 | |
| 682 | |
| 683 | def _seed_oss_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 684 | parts = [f"\n<function={tc.name}>"] |
| 685 | for key, value in tc.arguments.items(): |
| 686 | parts.append(f"\n<parameter={key}>{_qwen3_arg_value(value)}</parameter>") |
| 687 | parts.append("\n</function>\n") |
| 688 | return [ |
| 689 | ("<seed:tool_call>", True), |
| 690 | ("".join(parts), False), |
| 691 | ("</seed:tool_call>", True), |
| 692 | ] |
| 693 | |
| 694 | |
| 695 | def _seed_oss_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 696 | segs: list[tuple[str, bool]] = [] |
| 697 | if scenario.reasoning is not None: |
| 698 | segs.append((scenario.reasoning, False)) |
| 699 | if scenario.content is not None or scenario.tool_calls is not None: |
| 700 | segs.append(("</seed:think>", True)) |
| 701 | if scenario.tool_calls is not None and not scenario.tool_calls: |
| 702 | segs.append(("<seed:tool_call>", True)) |
| 703 | segs.append(("</seed:tool_call>", True)) |
| 704 | if scenario.content is not None: |
| 705 | segs.append((scenario.content, False)) |
| 706 | if scenario.tool_calls: |
| 707 | for tc in scenario.tool_calls: |
| 708 | segs.extend(_seed_oss_tool_segments(tc)) |
| 709 | return segs |
| 710 | |
| 711 | |
| 712 | def _build_seed_oss(scenario: Scenario, validate: bool = True) -> Sample: |
| 713 | sample = _make_sample( |
| 714 | sample_id=f"seed_oss-{scenario.id}", |
| 715 | description=scenario.description, |
| 716 | vocab=_SEED_OSS_VOCAB, |
| 717 | segments=_seed_oss_segments(scenario), |
| 718 | expected_reasoning=scenario.reasoning if scenario.reasoning is not None else "", |
| 719 | expected_content=_qwen3_expected_content(scenario), |
| 720 | expected_tool_calls=_expected_tc(scenario), |
| 721 | tools=_expected_tools(scenario), |
| 722 | ) |
| 723 | if validate: |
| 724 | _validate_sample(sample, SeedOssParser) |
| 725 | return sample |
| 726 | |
| 727 | |
| 728 | # ── DeepSeek V4 (DSML tool format) ────────────────────────────────── |
| 729 | |
| 730 | _DSML = "｜DSML｜" |
| 731 | _DSV4_VOCAB: dict[str, int] = { |
| 732 | "<think>": 128821, |
| 733 | "</think>": 128822, |
| 734 | f"<{_DSML}tool_calls>": 128823, |
| 735 | f"</{_DSML}tool_calls>": 128824, |
| 736 | } |
| 737 | |
| 738 | |
| 739 | def _dsv4_param_text(key: str, value: Any, dsml: str = _DSML) -> str: |
| 740 | is_string = isinstance(value, str) |
| 741 | if is_string: |
| 742 | val_str = value |
| 743 | elif isinstance(value, bool): |
| 744 | val_str = "true" if value else "false" |
| 745 | elif isinstance(value, (int, float)): |
| 746 | val_str = str(value) |
| 747 | else: |
| 748 | val_str = json.dumps(value, ensure_ascii=False) |
| 749 | string_attr = "true" if is_string else "false" |
| 750 | return ( |
| 751 | f'<{dsml}parameter name="{key}" string="{string_attr}">' |
| 752 | f"{val_str}</{dsml}parameter>\n" |
| 753 | ) |
| 754 | |
| 755 | |
| 756 | def _dsv4_tool_text(tc: ToolCallSpec, dsml: str = _DSML) -> str: |
| 757 | parts = [f'<{dsml}invoke name="{tc.name}">\n'] |
| 758 | for key, value in tc.arguments.items(): |
| 759 | parts.append(_dsv4_param_text(key, value, dsml)) |
| 760 | parts.append(f"</{dsml}invoke>\n") |
| 761 | return "".join(parts) |
| 762 | |
| 763 | |
| 764 | def _dsml_tool_segs( |
| 765 | scenario: Scenario, |
| 766 | tag: str, |
| 767 | dsml: str = _DSML, |
| 768 | ) -> list[tuple[str, bool]]: |
| 769 | if not scenario.tool_calls: |
| 770 | return [] |
| 771 | parts = ["\n"] |
| 772 | for tc in scenario.tool_calls: |
| 773 | parts.append(_dsv4_tool_text(tc, dsml)) |
| 774 | return [ |
| 775 | (f"<{dsml}{tag}>", True), |
| 776 | ("".join(parts), False), |
| 777 | (f"</{dsml}{tag}>", True), |
| 778 | ] |
| 779 | |
| 780 | |
| 781 | def _dsv4_segments( |
| 782 | scenario: Scenario, thinking: bool, *, v41: bool = False |
| 783 | ) -> list[tuple[str, bool]]: |
| 784 | segs: list[tuple[str, bool]] = [] |
| 785 | |
| 786 | if thinking: |
| 787 | if scenario.reasoning is not None: |
| 788 | segs.append((scenario.reasoning, False)) |
| 789 | if scenario.content is not None or scenario.tool_calls: |
| 790 | segs.append(("</think>", True)) |
| 791 | else: |
| 792 | if scenario.reasoning is not None: |
| 793 | segs.append(("<think>", True)) |
| 794 | segs.append((scenario.reasoning, False)) |
| 795 | segs.append(("</think>", True)) |
| 796 | |
| 797 | if scenario.content is not None: |
| 798 | segs.append((scenario.content, False)) |
| 799 | |
| 800 | segs.extend( |
| 801 | _dsml_tool_segs(scenario, "calls", f"{_DSML} ") |
| 802 | if v41 |
| 803 | else _dsml_tool_segs(scenario, "tool_calls") |
| 804 | ) |
| 805 | return segs |
| 806 | |
| 807 | |
| 808 | def _build_deepseek_v4( |
| 809 | scenario: Scenario, validate: bool = True, *, v41: bool = False |
| 810 | ) -> Sample: |
| 811 | thinking = scenario.reasoning is not None |
| 812 | chat_kwargs = {"thinking": thinking} |
| 813 | parser_cls = DeepSeekV41Parser if v41 else DeepSeekV4Parser |
| 814 | model = "deepseek_v41" if v41 else "deepseek_v4" |
| 815 | vocab = ( |
| 816 | { |
| 817 | "<think>": 128821, |
| 818 | "</think>": 128822, |
| 819 | f"<{_DSML} calls>": 128823, |
| 820 | f"</{_DSML} calls>": 128824, |
| 821 | } |
| 822 | if v41 |
| 823 | else _DSV4_VOCAB |
| 824 | ) |
| 825 | |
| 826 | if thinking: |
| 827 | expected_reasoning: str | None = scenario.reasoning or "" |
| 828 | else: |
| 829 | expected_reasoning = None |
| 830 | |
| 831 | sample = _make_sample( |
| 832 | sample_id=f"{model}-{scenario.id}", |
| 833 | description=scenario.description, |
| 834 | vocab=vocab, |
| 835 | segments=_dsv4_segments(scenario, thinking, v41=v41), |
| 836 | expected_reasoning=expected_reasoning, |
| 837 | expected_content=_qwen3_expected_content(scenario), |
| 838 | expected_tool_calls=_expected_tc(scenario), |
| 839 | tools=_expected_tools(scenario), |
| 840 | chat_template_kwargs=chat_kwargs, |
| 841 | ) |
| 842 | if validate: |
| 843 | kwargs = {} |
| 844 | if chat_kwargs: |
| 845 | kwargs["chat_template_kwargs"] = chat_kwargs |
| 846 | _validate_sample(sample, parser_cls, **kwargs) |
| 847 | return sample |
| 848 | |
| 849 | |
| 850 | # ── DeepSeek V3.2 (DSML tool format, no reasoning) ────────────────── |
| 851 | |
| 852 | _DSV32_VOCAB: dict[str, int] = { |
| 853 | f"<{_DSML}function_calls>": 128830, |
| 854 | f"</{_DSML}function_calls>": 128831, |
| 855 | } |
| 856 | |
| 857 | |
| 858 | def _dsv32_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 859 | segs: list[tuple[str, bool]] = [] |
| 860 | |
| 861 | if scenario.content is not None: |
| 862 | segs.append((scenario.content, False)) |
| 863 | |
| 864 | segs.extend(_dsml_tool_segs(scenario, "function_calls")) |
| 865 | return segs |
| 866 | |
| 867 | |
| 868 | def _build_deepseek_v32(scenario: Scenario, validate: bool = True) -> Sample | None: |
| 869 | if scenario.reasoning is not None: |
| 870 | return None |
| 871 | |
| 872 | sample = _make_sample( |
| 873 | sample_id=f"deepseek_v32-{scenario.id}", |
| 874 | description=scenario.description, |
| 875 | vocab=_DSV32_VOCAB, |
| 876 | segments=_dsv32_segments(scenario), |
| 877 | expected_reasoning=None, |
| 878 | expected_content=_qwen3_expected_content(scenario), |
| 879 | expected_tool_calls=_expected_tc(scenario), |
| 880 | tools=_expected_tools(scenario), |
| 881 | ) |
| 882 | if validate: |
| 883 | _validate_sample(sample, DeepSeekV32Parser) |
| 884 | return sample |
| 885 | |
| 886 | |
| 887 | # ── GLM-4.7 MoE (XML tool format, starts in REASONING) ────────────── |
| 888 | |
| 889 | _GLM47_MOE_VOCAB: dict[str, int] = { |
| 890 | "<think>": 50, |
| 891 | "</think>": 51, |
| 892 | "<tool_call>": 60, |
| 893 | "</tool_call>": 61, |
| 894 | "<arg_key>": 62, |
| 895 | "</arg_key>": 63, |
| 896 | "<arg_value>": 64, |
| 897 | "</arg_value>": 65, |
| 898 | } |
| 899 | |
| 900 | |
| 901 | def _glm47_moe_arg_value(value: Any) -> str: |
| 902 | if isinstance(value, bool): |
| 903 | return "true" if value else "false" |
| 904 | if isinstance(value, (int, float)): |
| 905 | return str(value) |
| 906 | if isinstance(value, str): |
| 907 | return value |
| 908 | return json.dumps(value, ensure_ascii=False) |
| 909 | |
| 910 | |
| 911 | def _glm47_moe_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 912 | segs: list[tuple[str, bool]] = [ |
| 913 | ("<tool_call>", True), |
| 914 | (tc.name, False), |
| 915 | ] |
| 916 | for key, value in tc.arguments.items(): |
| 917 | segs.extend( |
| 918 | [ |
| 919 | ("<arg_key>", True), |
| 920 | (key, False), |
| 921 | ("</arg_key>", True), |
| 922 | ("<arg_value>", True), |
| 923 | (_glm47_moe_arg_value(value), False), |
| 924 | ("</arg_value>", True), |
| 925 | ] |
| 926 | ) |
| 927 | segs.append(("</tool_call>", True)) |
| 928 | return segs |
| 929 | |
| 930 | |
| 931 | def _glm47_moe_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 932 | segs: list[tuple[str, bool]] = [] |
| 933 | if scenario.reasoning is not None: |
| 934 | segs.append((scenario.reasoning, False)) |
| 935 | if scenario.content is not None or scenario.tool_calls: |
| 936 | segs.append(("</think>", True)) |
| 937 | if scenario.content is not None: |
| 938 | segs.append((scenario.content, False)) |
| 939 | if scenario.tool_calls: |
| 940 | for tc in scenario.tool_calls: |
| 941 | segs.extend(_glm47_moe_tool_segments(tc)) |
| 942 | return segs |
| 943 | |
| 944 | |
| 945 | def _build_glm47_moe(scenario: Scenario, validate: bool = True) -> Sample: |
| 946 | sample = _make_sample( |
| 947 | sample_id=f"glm47_moe-{scenario.id}", |
| 948 | description=scenario.description, |
| 949 | vocab=_GLM47_MOE_VOCAB, |
| 950 | segments=_glm47_moe_segments(scenario), |
| 951 | expected_reasoning=scenario.reasoning if scenario.reasoning is not None else "", |
| 952 | expected_content=_qwen3_expected_content(scenario), |
| 953 | expected_tool_calls=_expected_tc(scenario), |
| 954 | tools=_expected_tools(scenario), |
| 955 | ) |
| 956 | if validate: |
| 957 | _validate_sample(sample, Glm47MoeParser) |
| 958 | return sample |
| 959 | |
| 960 | |
| 961 | # ── Kimi K2 (native tool-call section, starts in REASONING) ────────── |
| 962 | |
| 963 | _KIMI_K2_VOCAB: dict[str, int] = { |
| 964 | "<think>": 50, |
| 965 | "</think>": 51, |
| 966 | "<|tool_calls_section_begin|>": 60, |
| 967 | "<|tool_calls_section_end|>": 61, |
| 968 | "<|tool_call_begin|>": 62, |
| 969 | "<|tool_call_end|>": 63, |
| 970 | "<|tool_call_argument_begin|>": 64, |
| 971 | } |
| 972 | |
| 973 | |
| 974 | def _kimi_k2_tool_segments( |
| 975 | tool_calls: list[ToolCallSpec], |
| 976 | ) -> list[tuple[str, bool]]: |
| 977 | segs: list[tuple[str, bool]] = [("<|tool_calls_section_begin|>", True)] |
| 978 | for index, tc in enumerate(tool_calls): |
| 979 | args = json.dumps(tc.arguments, ensure_ascii=False, separators=(",", ":")) |
| 980 | segs.extend( |
| 981 | [ |
| 982 | ("<|tool_call_begin|>", True), |
| 983 | (f"functions.{tc.name}:{index}\n", False), |
| 984 | ("<|tool_call_argument_begin|>", True), |
| 985 | (args, False), |
| 986 | ("<|tool_call_end|>", True), |
| 987 | ] |
| 988 | ) |
| 989 | segs.append(("<|tool_calls_section_end|>", True)) |
| 990 | return segs |
| 991 | |
| 992 | |
| 993 | def _kimi_k2_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 994 | segs: list[tuple[str, bool]] = [] |
| 995 | if scenario.reasoning is not None: |
| 996 | segs.append(("<think>", True)) |
| 997 | segs.append((scenario.reasoning, False)) |
| 998 | if scenario.content is not None or scenario.tool_calls is not None: |
| 999 | segs.append(("</think>", True)) |
| 1000 | if scenario.content is not None: |
| 1001 | segs.append((scenario.content, False)) |
| 1002 | if scenario.tool_calls is not None: |
| 1003 | segs.extend(_kimi_k2_tool_segments(scenario.tool_calls)) |
| 1004 | return segs |
| 1005 | |
| 1006 | |
| 1007 | def _build_kimi_k2( |
| 1008 | scenario: Scenario, |
| 1009 | validate: bool = True, |
| 1010 | thinking: bool = True, |
| 1011 | ) -> Sample: |
| 1012 | expected_reasoning = ( |
| 1013 | scenario.reasoning.rstrip() |
| 1014 | if (thinking and scenario.reasoning is not None) |
| 1015 | else None |
| 1016 | ) |
| 1017 | if thinking and scenario.reasoning is None: |
| 1018 | expected_reasoning = "" |
| 1019 | |
| 1020 | sample = _make_sample( |
| 1021 | sample_id=f"kimi_k2-{scenario.id}", |
| 1022 | description=scenario.description, |
| 1023 | vocab=_KIMI_K2_VOCAB, |
| 1024 | segments=_kimi_k2_segments(scenario), |
| 1025 | expected_reasoning=expected_reasoning, |
| 1026 | expected_content=_qwen3_expected_content(scenario), |
| 1027 | expected_tool_calls=_expected_tc(scenario), |
| 1028 | tools=_expected_tools(scenario), |
| 1029 | chat_template_kwargs=None if thinking else {"thinking": False}, |
| 1030 | ) |
| 1031 | if validate: |
| 1032 | _validate_sample( |
| 1033 | sample, |
| 1034 | KimiK2Parser, |
| 1035 | chat_template_kwargs=sample.chat_template_kwargs, |
| 1036 | ) |
| 1037 | return sample |
| 1038 | |
| 1039 | |
| 1040 | _KIMI_K2_SCENARIOS = [ |
| 1041 | *SCENARIOS, |
| 1042 | Scenario( |
| 1043 | id="trailing-reasoning-whitespace", |
| 1044 | description="Reasoning trailing whitespace is stripped", |
| 1045 | reasoning="Reasoning with trailing whitespace. \n\t", |
| 1046 | content="Done.", |
| 1047 | ), |
| 1048 | ] |
| 1049 | |
| 1050 | |
| 1051 | # ── Inkling (typed content blocks, JSON tool payloads) ─────────────────── |
| 1052 | |
| 1053 | _TML_VOCAB: dict[str, int] = { |
| 1054 | "<|message_model|>": 200001, |
| 1055 | "<|content_text|>": 200004, |
| 1056 | "<|content_model_end_sampling|>": 200006, |
| 1057 | "<|content_thinking|>": 200008, |
| 1058 | "<|end_message|>": 200010, |
| 1059 | "<|content_tool_error|>": 200022, |
| 1060 | "<|content_invoke_tool_json|>": 200049, |
| 1061 | "<|content_invoke_tool_text|>": 200057, |
| 1062 | } |
| 1063 | |
| 1064 | |
| 1065 | def _inkling_block( |
| 1066 | segs: list[tuple[str, bool]], |
| 1067 | kind_token: str, |
| 1068 | body: str, |
| 1069 | ) -> None: |
| 1070 | """Append one Inkling content block; blocks after the first start with |
| 1071 | the ``<|message_model|>`` role token (the first block continues the |
| 1072 | generation prompt directly).""" |
| 1073 | if segs: |
| 1074 | segs.append(("<|message_model|>", True)) |
| 1075 | segs.append((kind_token, True)) |
| 1076 | if body: |
| 1077 | segs.append((body, False)) |
| 1078 | segs.append(("<|end_message|>", True)) |
| 1079 | |
| 1080 | |
| 1081 | def _inkling_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 1082 | segs: list[tuple[str, bool]] = [] |
| 1083 | if scenario.reasoning is not None: |
| 1084 | _inkling_block(segs, "<|content_thinking|>", scenario.reasoning) |
| 1085 | if scenario.tool_calls is not None and not scenario.tool_calls: |
| 1086 | _inkling_block(segs, "<|content_invoke_tool_json|>", "") |
| 1087 | if scenario.content is not None: |
| 1088 | _inkling_block(segs, "<|content_text|>", scenario.content) |
| 1089 | if scenario.tool_calls: |
| 1090 | for tc in scenario.tool_calls: |
| 1091 | args = json.dumps(tc.arguments, ensure_ascii=False, separators=(",", ":")) |
| 1092 | payload = f'{{"name":"{tc.name}","args":{args}}}' |
| 1093 | _inkling_block(segs, "<|content_invoke_tool_json|>", payload) |
| 1094 | return segs |
| 1095 | |
| 1096 | |
| 1097 | def _build_inkling(scenario: Scenario, validate: bool = True) -> Sample: |
| 1098 | prompt_token_ids = None |
| 1099 | if scenario.after_tool_response: |
| 1100 | # Prompt ends with a closed tool-response block and the |
| 1101 | # generation-prompt role token. |
| 1102 | prompt_token_ids = [ |
| 1103 | _TML_VOCAB["<|end_message|>"], |
| 1104 | _TML_VOCAB["<|message_model|>"], |
| 1105 | ] |
| 1106 | sample = _make_sample( |
| 1107 | sample_id=f"inkling-{scenario.id}", |
| 1108 | description=scenario.description, |
| 1109 | vocab=_TML_VOCAB, |
| 1110 | segments=_inkling_segments(scenario), |
| 1111 | expected_reasoning=scenario.reasoning, |
| 1112 | expected_content=_qwen3_expected_content(scenario), |
| 1113 | expected_tool_calls=_expected_tc(scenario), |
| 1114 | tools=_expected_tools(scenario), |
| 1115 | prompt_token_ids=prompt_token_ids, |
| 1116 | ) |
| 1117 | if validate: |
| 1118 | _validate_sample(sample, InklingParser) |
| 1119 | return sample |
| 1120 | |
| 1121 | |
| 1122 | # ── PLaMo3 (literal nested tool-request wrappers) ─────────────────── |
| 1123 | |
| 1124 | _PLAMO3_MARKERS = ( |
| 1125 | plamo3.BEGIN_THINK, |
| 1126 | plamo3.END_THINK, |
| 1127 | plamo3.BEGIN_TOOL_REQUESTS, |
| 1128 | plamo3.END_TOOL_REQUESTS, |
| 1129 | plamo3.BEGIN_TOOL_REQUEST, |
| 1130 | plamo3.END_TOOL_REQUEST, |
| 1131 | plamo3.BEGIN_TOOL_NAME, |
| 1132 | plamo3.END_TOOL_NAME, |
| 1133 | plamo3.BEGIN_TOOL_ARGUMENTS, |
| 1134 | plamo3.END_TOOL_ARGUMENTS, |
| 1135 | plamo3.EOT, |
| 1136 | ) |
| 1137 | _PLAMO3_VOCAB = {marker: 300 + i for i, marker in enumerate(_PLAMO3_MARKERS)} |
| 1138 | |
| 1139 | |
| 1140 | def _plamo3_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]: |
| 1141 | args = json.dumps(tc.arguments, ensure_ascii=False, separators=(",", ":")) |
| 1142 | return [ |
| 1143 | (plamo3.BEGIN_TOOL_REQUEST, True), |
| 1144 | (plamo3.BEGIN_TOOL_NAME, True), |
| 1145 | (tc.name, False), |
| 1146 | (plamo3.END_TOOL_NAME, True), |
| 1147 | (plamo3.BEGIN_TOOL_ARGUMENTS, True), |
| 1148 | (args, False), |
| 1149 | (plamo3.END_TOOL_ARGUMENTS, True), |
| 1150 | (plamo3.END_TOOL_REQUEST, True), |
| 1151 | ] |
| 1152 | |
| 1153 | |
| 1154 | def _plamo3_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 1155 | segs: list[tuple[str, bool]] = [ |
| 1156 | (plamo3.BEGIN_THINK, True), |
| 1157 | (scenario.reasoning or "", False), |
| 1158 | (plamo3.END_THINK, True), |
| 1159 | ] |
| 1160 | if scenario.content is not None: |
| 1161 | segs.append((scenario.content, False)) |
| 1162 | if scenario.tool_calls is not None: |
| 1163 | segs.append((plamo3.BEGIN_TOOL_REQUESTS, True)) |
| 1164 | for tool_call in scenario.tool_calls: |
| 1165 | segs.extend(_plamo3_tool_segments(tool_call)) |
| 1166 | segs.append((plamo3.END_TOOL_REQUESTS, True)) |
| 1167 | return segs |
| 1168 | |
| 1169 | |
| 1170 | def _build_plamo3(scenario: Scenario, validate: bool = True) -> Sample: |
| 1171 | sample = _make_sample( |
| 1172 | sample_id=f"plamo3-{scenario.id}", |
| 1173 | description=scenario.description, |
| 1174 | vocab=_PLAMO3_VOCAB, |
| 1175 | segments=_plamo3_segments(scenario), |
| 1176 | expected_reasoning=scenario.reasoning or "", |
| 1177 | expected_content=_qwen3_expected_content(scenario), |
| 1178 | expected_tool_calls=_expected_tc(scenario), |
| 1179 | tools=_expected_tools(scenario), |
| 1180 | ) |
| 1181 | if validate: |
| 1182 | _validate_sample(sample, Plamo3Parser) |
| 1183 | return sample |
| 1184 | |
| 1185 | |
| 1186 | # ── Granite (JSON-array tool bodies, no reasoning) ─────────────────────── |
| 1187 | |
| 1188 | _GRANITE_VOCAB: dict[str, int] = { |
| 1189 | "<|tool_call|>": 49154, |
| 1190 | } |
| 1191 | |
| 1192 | |
| 1193 | def _granite_segments(scenario: Scenario) -> list[tuple[str, bool]]: |
| 1194 | segs: list[tuple[str, bool]] = [] |
| 1195 | if scenario.content: |
| 1196 | # Granite has no reasoning; prose is plain content preceding the marker. |
| 1197 | segs.append((scenario.content, False)) |
| 1198 | if scenario.tool_calls: |
| 1199 | segs.append(("<|tool_call|>", True)) |
| 1200 | payload = json.dumps( |
| 1201 | [ |
| 1202 | {"name": tc.name, "arguments": tc.arguments} |
| 1203 | for tc in scenario.tool_calls |
| 1204 | ], |
| 1205 | ensure_ascii=False, |
| 1206 | separators=(", ", ": "), |
| 1207 | ) |
| 1208 | segs.append((" " + payload, False)) |
| 1209 | return segs |
| 1210 | |
| 1211 | |
| 1212 | def _granite_expected_content(scenario: Scenario) -> str | None: |
| 1213 | if scenario.tool_calls: |
| 1214 | if not scenario.content: |
| 1215 | return None |
| 1216 | return scenario.content.strip() or None |
| 1217 | return scenario.content |
| 1218 | |
| 1219 | |
| 1220 | def _build_granite(scenario: Scenario, validate: bool = True) -> Sample: |
| 1221 | sample = _make_sample( |
| 1222 | sample_id=f"granite-{scenario.id}", |
| 1223 | description=scenario.description, |
| 1224 | vocab=_GRANITE_VOCAB, |
| 1225 | segments=_granite_segments(scenario), |
| 1226 | expected_reasoning=None, |
| 1227 | expected_content=_granite_expected_content(scenario), |
| 1228 | expected_tool_calls=_expected_tc(scenario), |
| 1229 | tools=_expected_tools(scenario), |
| 1230 | ) |
| 1231 | if validate: |
| 1232 | _validate_sample(sample, GraniteParser) |
| 1233 | return sample |
| 1234 | |
| 1235 | |
| 1236 | # Granite has no reasoning, so the shared reasoning-centric SCENARIOS do not |
| 1237 | # apply; these exercise the JSON-array tool body (single, parallel, surrounding |
| 1238 | # text) instead. |
| 1239 | _GRANITE_SCENARIOS: list[Scenario] = [ |
| 1240 | Scenario( |
| 1241 | id="single-tool", |
| 1242 | description="Single tool call", |
| 1243 | tool_calls=[_READ_TOOL], |
| 1244 | ), |
| 1245 | Scenario( |
| 1246 | id="parallel-tools", |
| 1247 | description="Parallel tool calls in one JSON array", |
| 1248 | tool_calls=[_BASH_TOOL, _WEATHER_TOOL], |
| 1249 | ), |
| 1250 | Scenario( |
| 1251 | id="complex-json-args", |
| 1252 | description="Tool call with nested objects, arrays, numbers, booleans", |
| 1253 | tool_calls=[_COMPLEX_TOOL], |
| 1254 | ), |
| 1255 | Scenario( |
| 1256 | id="content-only", |
| 1257 | description="Plain content response without tool calls", |
| 1258 | content="Hello! How can I help you today?", |
| 1259 | ), |
| 1260 | Scenario( |
| 1261 | id="surrounding-text", |
| 1262 | description="Prose content preceding the tool call", |
| 1263 | content="Let me check the weather.", |
| 1264 | tool_calls=[_WEATHER_TOOL], |
| 1265 | ), |
| 1266 | ] |
| 1267 | |
| 1268 | |
| 1269 | # ── Registry and public API ────────────────────────────────────────── |
| 1270 | |
| 1271 | _BUILDERS: dict[str, Any] = { |
| 1272 | "deepseek_v32": _build_deepseek_v32, |
| 1273 | "deepseek_v4": _build_deepseek_v4, |
| 1274 | "deepseek_v41": functools.partial(_build_deepseek_v4, v41=True), |
| 1275 | "gemma4": _build_gemma4, |
| 1276 | "granite": _build_granite, |
| 1277 | "minimax_m2": _build_minimax_m2, |
| 1278 | "nemotron_v3": _build_nemotron_v3, |
| 1279 | "granite_thinking_parser": _build_granite_thinking, |
| 1280 | "seed_oss": _build_seed_oss, |
| 1281 | "glm47_moe": _build_glm47_moe, |
| 1282 | "kimi_k2": _build_kimi_k2, |
| 1283 | "qwen3": _build_qwen3, |
| 1284 | "mimo": _build_mimo, |
| 1285 | "step3p5": _build_step3p5, |
| 1286 | "inkling": _build_inkling, |
| 1287 | "plamo3": _build_plamo3, |
| 1288 | } |
| 1289 | |
| 1290 | |
| 1291 | @functools.cache |
| 1292 | def build_samples(model: str) -> tuple[Sample, ...]: |
| 1293 | """Build all scenario samples for a model, self-validated.""" |
| 1294 | builder = _BUILDERS[model] |
| 1295 | scenarios = _KIMI_K2_SCENARIOS if model == "kimi_k2" else SCENARIOS |
| 1296 | return tuple(s for s in (builder(sc) for sc in scenarios) if s is not None) |