Predibase

Predibase provides a platform for developers to fine-tune and serve large language models (LLMs) in a fast and economical way. Created by AI experts from Uber, Google, Apple and Amazon, Predibase streamlines fine-tuning open-source LLMs like Llama-3, CodeLlama and Mistral for a particular job.

Predibase lets you fine-tune any open-source LLM for your particular use case, such as classification, information extraction, customer sentiment analysis, code generation and named entity recognition. The platform offers a range of features to make fine-tuning more efficient and economical:

• First-class fine-tuning experience: Supports state-of-the-art methods like quantization, low-rank adaptation and memory-efficient distributed training for fast and efficient fine-tuning jobs.
• Cost-effective serving infrastructure: LoRAX (LoRA eXchange) technology enables serving many fine-tuned adapters on a single GPU, dramatically lowering costs.
• Free serverless inference: Up to 1 million tokens per day and 10 million tokens per month for prototyping, evaluation and experimental use cases.
• Enterprise-grade security: Built with enterprise-grade security and SOC-2 compliance, so users can download and export trained models at any time.

With Predibase, developers can rapidly deploy and query open-source pre-trained LLMs like Llama-2, Mistral and Zephyr, experiment and evaluate the best base model for their use case, and fine-tune models for specific tasks with ease. The platform also supports a wide range of models, including BioMistral, CodeLlama, Gemma, GPT-2, Llama 2, Llama 3, Meditron, Mixtral and Phi.

Predibase uses a pay-as-you-go pricing model for its Developer tier, with prices that depend on the size of the model being fine-tuned and the size of the dataset used. For example, fine-tuning a model up to 7B parameters costs $0.36 per 1 million tokens. Dedicated deployments are also available with usage-based pricing billed by the second, offering high inference performance and automatic autoscaling.

Although designed for developers, Predibase also supports Enterprise and Virtual Private Cloud (VPC) deployments, offering dedicated support channels and priority access to A100 GPUs. This platform is great for organizations that want to fine-tune and serve LLMs for specific use cases while keeping control over their models and data.

Predibase is a useful option for those looking for a developer-focused platform to fine-tune and serve large language models, making it a good fit for a broad range of AI use cases.

Published on June 13, 2024