| """Task-agnostic decision interfaces. |
| |
| MiniJEV always answers the same kind of request:: |
| |
| DecisionRequest(state=..., question="...", choices=["a", "b", ...]) -> Decision |
| |
| `state` is opaque to the core: each environment/task plugin knows how to turn |
| its states into tensors (see `minijev.tasks.Task`). Choices are plain strings, |
| so the number and the meaning of choices are *inputs*, not model constants. |
| """ |
| from __future__ import annotations |
|
|
| from dataclasses import dataclass, field |
| from typing import Any, Sequence |
|
|
| State = Any |
|
|
|
|
| @dataclass |
| class DecisionRequest: |
| state: State |
| question: str |
| choices: Sequence[str] |
|
|
|
|
| @dataclass |
| class Decision: |
| choice: str |
| index: int |
| confidence: float |
| scores: dict[str, float] |
| logits: list[float] = field(default_factory=list) |
| embedding: list[float] | None = None |
| raw_scores: dict[str, float] | None = None |
| temperature: float = 1.0 |
|
|
| @property |
| def distribution(self) -> dict[str, float]: |
| return self.scores |
|
|
| @property |
| def raw_confidence(self) -> float: |
| return self.raw_scores[self.choice] if self.raw_scores else self.confidence |
|
|
| def to_dict(self) -> dict: |
| d = {"choice": self.choice, "confidence": round(self.confidence, 4), |
| "distribution": {k: round(v, 4) for k, v in self.scores.items()}} |
| if self.raw_scores is not None: |
| d["raw_distribution"] = {k: round(v, 4) for k, v in self.raw_scores.items()} |
| if self.embedding is not None: |
| d["embedding"] = self.embedding |
| return d |
|
|
|
|
| @dataclass |
| class BooleanDecision: |
| answer: bool |
| confidence: float |
| p_yes: float |
| raw_p_yes: float |
|
|
| def __bool__(self) -> bool: |
| return self.answer |
|
|
