Stop fine-tuning open-source models on garbage data pipelines

Every week, a team drops considerable engineering cycles into fine-tuning the latest open-source weight release, only to watch their model hallucinate syntax, ignore system prompts, and fail basic reg