Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts