Refine inference instructions for concise segments

inference.py

@@ -22,7 +22,7 @@ from transformers import Qwen2_5OmniProcessor
 
 ADAPTER_PATH = (Path(__file__).resolve().parent / "model" / "checkpoint-22000").resolve()
 DEFAULT_SYSTEM_PROMPT = "You are an audio segmentation assistant. For each turn, return JSON with the requested keys."
-DEFAULT_INSTRUCTION = "Provide a JSON object with start_time, end_time, and caption fields describing the first salient event in the clip."
+DEFAULT_INSTRUCTION = "General note: The segments should be relatively concise (1-8 seconds). To begin, respond with a JSON object containing `start_time` and `end_time` for the initial segment. Return the JSON only, with no extra commentary."
 USE_AUDIO_IN_VIDEO = False