Spaces:
Running on Zero
Running on Zero
| """The playground's task presets: prompts, the JSON schema each can be constrained to, and | |
| whether the schema is on by default. | |
| Emotion the seven-label emotion prompt, no system prompt of our own; free output by default | |
| Combined the Inter-2 Audio system prompt + one instruction covering engagement, social | |
| signals, emotion and delivery; constrained by default, since free output can skip | |
| sections | |
| When constrained, every label is limited to the values listed in the prompt and every | |
| rationale is capped at 300 characters. | |
| """ | |
| from __future__ import annotations | |
| from dataclasses import dataclass | |
| from pathlib import Path | |
| PROMPTS = Path(__file__).parent / "prompts" | |
| EMOTIONS = ["angry", "disgusted", "fearful", "happy", "neutral", "sad", "surprised"] | |
| CATALOG_EMOTIONS = ["Anger", "Disgust", "Fear", "Happiness", "Neutral", "Sadness", "Surprise"] | |
| SIGNALS = ["Hesitation", "Agreement", "Uncertainty", "Frustration", "Interest", "Confidence"] | |
| ENGAGEMENT = ["engaged", "neutral", "disengaged"] | |
| DELIVERY = ["Normal", "Crying", "Laughing", "Mumbling", "Shouting", "Singing", "Whispering"] | |
| LEVELS = ["high", "medium", "low"] | |
| RATIONALE_MAX_CHARS = 300 # keeps the longest constrained answer well inside 768 new tokens | |
| def _prompt(name: str) -> str: | |
| return (PROMPTS / name).read_text(encoding="utf-8").strip() | |
| def emotion_schema(confidence_steps: int = 4, evidence_max_chars: int = 0) -> dict: | |
| """{primary_emotion, confidence, evidence}: the label limited to the 7 emotions. | |
| confidence is limited to i/confidence_steps (a power of 2, so xgrammar keeps | |
| the short float form) and evidence is capped at evidence_max_chars (0 forces | |
| it empty): the label is what matters for this task, and an unbounded number | |
| can loop on digits. Properties are generated in this order, | |
| so the label comes first and neither cap can change it under greedy decoding. | |
| """ | |
| if confidence_steps < 1 or confidence_steps & (confidence_steps - 1): | |
| raise ValueError(f"confidence_steps must be a power of 2, got {confidence_steps}") | |
| return { | |
| "type": "object", | |
| "properties": { | |
| "primary_emotion": {"type": "string", "enum": EMOTIONS}, | |
| "confidence": {"enum": [i / confidence_steps for i in range(confidence_steps + 1)]}, | |
| "evidence": {"type": "string", "maxLength": evidence_max_chars}, | |
| }, | |
| "required": ["primary_emotion", "confidence", "evidence"], | |
| } | |
| def _labelled_list(names: list[str], min_items: int, max_items: int, rationale: bool = True) -> dict: | |
| fields = { | |
| "name": {"type": "string", "enum": names}, | |
| "probability": {"type": "string", "enum": LEVELS}, | |
| } | |
| if rationale: | |
| fields["rationale"] = {"type": "string", "maxLength": RATIONALE_MAX_CHARS} | |
| return { | |
| "type": "array", | |
| "minItems": min_items, | |
| "maxItems": max_items, | |
| "items": { | |
| "type": "object", | |
| "properties": fields, | |
| "required": list(fields), | |
| }, | |
| } | |
| def combined_schema() -> dict: | |
| """{engagement_status, signals, emotions, delivery}, in that order (the instruction's order).""" | |
| return { | |
| "type": "object", | |
| "properties": { | |
| "engagement_status": {"type": "string", "enum": ENGAGEMENT}, | |
| "signals": _labelled_list(SIGNALS, 0, len(SIGNALS)), | |
| "emotions": _labelled_list(CATALOG_EMOTIONS, 1, 1), | |
| "delivery": {"type": "string", "enum": DELIVERY}, | |
| }, | |
| "required": ["engagement_status", "signals", "emotions", "delivery"], | |
| } | |
| class Task: | |
| system_prompt: str | |
| user_prompt: str | |
| schema: dict | |
| constrain: bool # the schema checkbox's default for this task | |
| TASKS: dict[str, Task] = { | |
| "Emotion": Task("", _prompt("emotion_user.txt"), emotion_schema(), constrain=False), | |
| "Combined": Task(_prompt("shared_system.txt"), _prompt("combined_user.txt"), combined_schema(), constrain=True), | |
| } | |
| DEFAULT_TASK = "Emotion" | |