Curation for fine-tuning: tooling and common pitfalls

Discussion prompts: - Data sources: Docs, tickets, chats; how to de-duplicate and avoid leakage. - Quality control: Annotation guidelines, inter-annotator agreement. - Hard cases: Adversarial and corner cases; handling contradictions. - Balance: Not overfitting to happy paths; mix of easy and hard. - Tooling: Label Studio, Prodigy, or custom UIs; dataset registries. What acceptance gates do you enforce before training, and how do you keep datasets fresh without drift?