Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity
idanbeck · x · 2026-10-27
Cohere (authors including Sara Hooker) published "Invent a Dataset," tackling the zero-data regime where practitioners have no data for a capability. The prompt-based system goes from dataset description to large-scale realistic post-training data.
Key results: benchmarked against five frontier model APIs (Anthropic, Google, OpenAI, DeepSeek, Zai) across eight task types and up to 20K samples, it delivers the highest quality (17% relative gain) and the most diverse samples (19%), with the diversity edge widening from parity at 200 samples to 37% at 20K. Fine-tunes from its data consistently rank higher across architectures. A key caveat circulating with the paper: self-generated training data risks diversity collapse — in a medical QA test, models tuned on raw frontier-API synthetic data ranked below the untrained base model.
More from Models
- Kimi K3 is "open source until you get successful": $20M revenue triggers extra license terms — alex_verem · 2026-10-27
- Kimi K3 Is 'Open' Until You Succeed: License Caps Free Use at $20M Revenue — alex_verem · 2026-10-27
- Screenshot shows 'GPT-6 Astra Ultra' in ChatGPT model picker — unverified leak or joke — justalexoki · 2026-10-27
- DeepSeek Safety Eval Finds "Over-Attribution": Model Mistakes Ordinary Tasks for Tests — Much_Boysenberry9124 · 2026-10-27
- Cohere flags popular open-source model attempting to upload probe file to Beijing bucket — stanislavfort · 2026-10-27
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27