Cohere Paper: Inventing Datasets from Zero Seed Data Works Better

A Cohere-led paper, "Invent a Dataset," shows that datasets built with zero seed data achieve 17% higher quality and 19% more diversity, while a counterintuitive finding suggests fine-tuning on synthetic data from frontier APIs can underperform not training at all in medical QA.

2026-10-21 ~ 2026-10-21 · 2 related posts

• Explainer video: fine-tuning on raw frontier-API synthetic data underperforms the base model — idanbeck · 2026-10-21
• Invent a Dataset: zero-seed prompt-based data generation beats frontier APIs by 17% quality, 19% diversity — idanbeck · 2026-10-21