OpenPipe/mistral-ft-optimized-1218

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:8kPublished:Dec 17, 2023License:cc-by-nc-4.0Architecture:Transformer0.2K Open Weights Cold

OpenPipe/mistral-ft-optimized-1218 is a 7 billion parameter language model based on the Mistral-7B-v0.1 architecture, developed by OpenPipe. This model is a merge of Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp and Q-bert/MetaMath-Cybertron-Starling, optimized as a strong base for downstream fine-tuning across various tasks. It is designed to offer robust performance for custom applications, with a context length of 8192 tokens.

Loading preview...

OpenPipe/mistral-ft-optimized-1218 Overview

OpenPipe/mistral-ft-optimized-1218 is a 7 billion parameter language model built upon the Mistral-7B-v0.1 base. Developed by OpenPipe, this model is specifically engineered to serve as a highly capable foundation for subsequent fine-tuning across a diverse range of tasks. It leverages a sophisticated merge of two strong base models: Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp and Q-bert/MetaMath-Cybertron-Starling, utilizing the mergekit tool with a slerp merge method.

Key Characteristics

  • Optimized Base Model: Designed to be one of the strongest available base models for downstream fine-tuning, according to OpenPipe's internal evaluations.
  • Merge Architecture: Created by merging two high-performing models, combining their strengths for enhanced general utility.
  • Mistral-7B Foundation: Benefits from the efficient and capable architecture of Mistral-7B-v0.1.

Intended Use Cases

This model is primarily recommended for developers and researchers who require a robust and versatile base model to fine-tune for specific applications. Its design makes it suitable for:

  • Developing custom instruction-tuned models.
  • Adapting to niche domains or specialized tasks through further training.

Note: An updated version with similar performance and a more permissive license, OpenPipe/mistral-ft-optimized-1227, is available and recommended for most users.