Zephyr 7B Alpha
Hugging Face · 2023-10-09 · 7.2B parameters
Zephyr 7B Alpha is the first assistant-tuned model in Hugging Face H4's Zephyr series, produced by distilled supervised fine-tuning of Mistral 7B v0.1 on a filtered UltraChat corpus generated by GPT-3.5 Turbo, followed by distilled direct preference optimization on GPT-4-ranked UltraFeedback preferences. Released as an open research artifact, it demonstrated that a 7B model post-trained with public synthetic alignment data could compete with much larger chat models, while the authors cautioned that the work focused on helpfulness rather than safety alignment.