| """ |
| Canonical data models for Vox Dictation Refinement dataset. |
| Governed by GOAL.md and AGENTS.md. |
| Single-model non-autoregressive span extraction with single label: "speech disfluency". |
| """ |
|
|
| from typing import List, Literal, Optional, Tuple |
| from pydantic import BaseModel, Field |
|
|
|
|
| class RawCandidate(BaseModel): |
| """Raw output from Generator LLM: only the corrupted sentence and added disfluent spans.""" |
| raw_text: str = Field(description="The full corrupted spoken utterance") |
| clean_text: str = Field(description="The original clean sentence") |
| disfluent_spans: List[str] = Field( |
| default_factory=list, |
| description="Exact substrings added into raw_text (fillers, false starts, retractions) to excise" |
| ) |
|
|
|
|
| class Span(BaseModel): |
| """Character-level span representation with verified offsets.""" |
| label: Literal["speech disfluency"] = "speech disfluency" |
| span: Tuple[int, int] = Field(description="Exact [start, end] character offsets in raw_text") |
| text: str = Field(description="Exact substring in raw_text slice") |
| confidence: float = Field(default=1.0, ge=0.0, le=1.0) |
| origin: Optional[str] = Field( |
| default=None, |
| description=( |
| "PROVENANCE ONLY, never a training signal. Which gold annotation produced " |
| "this span: 'region:D' (disfluency), 'region:E' (edit/self-repair), " |
| "'region:F' (Tier-1 filled pause), 'correction' (the '[A, +B]' spurious->" |
| "repair group), 'synthetic:<archetype>' (LLM corruption), or None for " |
| "hand-written data. Used to stratify and audit the dataset mixture." |
| ), |
| ) |
|
|
|
|
| class DatasetExample(BaseModel): |
| """Canonical Unified JSONL Example for Training and Evaluation.""" |
| id: str |
| category: Literal["CLEAN", "DISFLUENCY"] |
| raw_text: str |
| clean_text: str |
| spans: List[Span] = Field(default_factory=list) |
|
|
|
|
| class AuditVerdict(BaseModel): |
| """Zero-temperature Judge audit result for an individual candidate.""" |
| id: str |
| verdict: Literal["PASS", "FAIL"] |
| category_accurate: bool |
| no_false_deletions: bool |
| number_entity_preserved: bool |
| reason: str |
|
|
