Thinking Machines Tinker

Tinker, by Thinking Machines Lab, is a powerful training API designed specifically for researchers and developers working with AI models. It simplifies the complexities of model training and fine-tuning by handling the infrastructure aspects.

Tinker provides users with complete control over their model's training and fine-tuning processes. This is achieved through four primary functions: forwardbackward for a forward and backward pass, optimstep for updating weights based on the accumulated gradient, sample for generating tokens for interaction, evaluation, or RL actions, and savestate for saving training progress.

Tinker further supports a wide range of open-source models, accommodating various needs and requirements of its users. To enhance efficiency, it utilizes LoRA, a method focused on training a small add-on instead of modifying all the original weights.

This streamlined approach matches the learning performance of full fine-tuning while offering more flexibility and requiring less compute. Tinker manages scheduling, tuning, resource management, and infrastructure reliability, which allows users to concentrate solely on their data and algorithms.

This also includes distributed training on powerful GPU clusters for effective utilization. No need for users to worry about hardware or infrastructure management.

Notably, Tinker maintains a strict privacy policy to ensure user data is exclusively used for fine-tuning their own models.