"""Minimal DOVE adapter template. Copy this file and replace ``create_predictor`` with your model loader. The returned callable receives one public question whose ``video`` is absolute. It must return only the target-interval caption as a string. """ from __future__ import annotations from typing import Any, Callable def create_predictor(model_args: dict[str, Any]) -> Callable[[dict[str, Any]], str]: # Load tokenizer/model once here. Values passed with --model-args are # available in model_args. del model_args def predict(question: dict[str, Any]) -> str: video_path = question["video"] prompt = question["prompt"] raise NotImplementedError( f"Run your model on video={video_path!r} with prompt={prompt!r}" ) return predict