Zephyr ORPO 141B

About

The Zephyr ORPO 141B is a cutting-edge large language model by Hugging Face, developed in partnership with Argilla and KAIST. It employs a Mixture of Experts (MoE) architecture, consisting of 141 billion parameters, with 39 billion active during operation. The model is derived from the Mixtral-8x22B framework and fine-tuned using the innovative Odds Ratio Preference Optimization (ORPO) method, which improves computational efficiency by removing the need for a separate supervised fine-tuning phase. Zephyr ORPO demonstrates impressive proficiency in tasks such as open-ended conversations, question answering, and coding assistance, evidenced by high scores on benchmarks like MT Bench (8.17) and IFEval (65.06) 139. The model, trained on a dataset of 7,000 instances, is optimized for multi-turn dialogues, making it ideal for interactive AI applications. However, attention should be paid to its lack of alignment with human safety preferences, as this could lead to problematic outputs 4512.

Capabilities

Providers(1)