Tailoring large language models (LLMs) to specific domains or use cases often requires significant computational resources and expertise. Mistral AI has introduced new tools and services that aim to make model customisation more efficient, cost-effective, and accessible to a wider range of developers and businesses.
On June 5, the Mistral AI team announced three different entry points for specialising their AI models:
1. Open-source fine-tuning SDK: Mistral-finetune is a lightweight and efficient codebase that allows developers to fine-tune Mistral's open-source models on their own infrastructure without sacrificing performance or memory efficiency. The codebase is built on the LoRA (Low-Rank Adaptation) training paradigm, which enables memory-efficient and performant fine-tuning.
2. Serverless fine-tuning services on la Plateforme: Mistral AI has introduced new fine-tuning services on their platform, which leverage the company's unique fine-tuning techniques to allow fast and cost-effective model adaptation and deployment. These services use LoRA adapters to prevent forgetting base model knowledge and enable efficient serving.
3. Custom training services: Mistral AI offers custom training services that involve fine-tuning their models on a customer's specific applications using their proprietary data. This approach enables the creation of highly specialised and optimised models for particular domains, using advanced techniques such as continuous pretraining to include proprietary knowledge within the model weights.
By making model customisation more accessible and cost-effective, businesses and developers can now theoretically create AI applications, tailored to their specific needs, without the need for extensive computational resources or expertise.