Research2026-08-10
Researchers at Hugging Face described Zephyr-7B, a small chat model trained to follow user instructions using rankings generated by another model rather than human labellers. The team reports it beat much larger open chat models on a common conversation benchmark and took only a few hours to train. Code, models and data were published openly.
What changed
Small models trained on outputs copied from larger models handled tasks well but responded poorly to ordinary conversational prompts, and improving that generally required human-labelled preference data.
What it unlocks
Training a small chat model to follow user intent using rankings produced by another model instead of human annotators, in a few hours.
- 7B parameters
- a few hours of training
- MT-Bench: surpasses Llama2-Chat-70B
What you need to act on it
- a ranked preference dataset produced by a teacher model
- arxiv.org2026-08-10