Glossary · Term

Zephyr

← all terms

Definition

Plain language

An open-weight Mistral-based assistant model fine-tuned with preference learning.

As stated in the literature

A Mistral-7B-based assistant model post-trained with DPO on preference data, used as an open-weights aligned model in alignment research.

Why it matters: It is a widely shared, reproducible reference point in open alignment work, making method comparisons easier across labs.

For example, a researcher comparing new alignment methods might use Zephyr as a strong, freely available baseline DPO-tuned model.

Heard on the show

“The dissociation rests on two natural-experiment pairs — Llama-three-point-one to three-point-three, and Mistral to Zephyr — plus the controlled DPO experiment.”
Episode 004 — The Sycophancy Circuit That Survives Alignment Training

Mentioned in 1 episode

  1. 004
    The Sycophancy Circuit That Survives Alignment Training

Related terms