The objective is straightforward: given a raw dialogue, the model produces a structured JSON object with topic and summary keys. This transforms unstructured conversational data into a machine-readable format suitable for downstream pipelines, databases, and APIs. Instruction-tuned models like Qwen2.5-3B-Instruct are capable of following JSON formatting instructions zero-shot, but they are inconsistent, output structure varies across prompts, edge cases produce malformed JSON, and there is no guarantee the keys will match a fixed schema. Fine-tuning on task-specific data solves this by teaching the model to reliably produce valid, schema-conformant JSON for any dialogue input. Demo: https://huggingface.co/spaces/llhax/dialog-to-json Code: https://github.com/HmadAfzal/qwen2.5-3B-Instruct-Dialog-to-json/