• Build Club
  • Posts
  • Build Memo Y24: Making customizing AI models faster and easier

Build Memo Y24: Making customizing AI models faster and easier

Build Memo: Build Club Accelerator Y24

Unsloth : Making customizing AI models faster and easier

πŸ› οΈ Make LLM training faster and finetuning 30x faster | πŸ‘©πŸ»β€πŸ’» Website

Hey, we are Daniel and Michael, the founders of Unsloth.

We are building Unsloth to make customizing AI models easier and faster.

πŸ€” The Problem

Customizing language models can take days or weeks to run, and is extremely hard and expensive to do.

πŸ’‘ The Solution

Put simply, Unsloth makes the whole process 2x faster and use 70% less memory, and much more accessible.

πŸ› οΈ How Do We Do It?

Our notebooks on our Github page are for free to use.

Simply upload your dataset you want to use to customize a language model like Llama, Mistral, Gemma or Phi, and press run all. You can export the customized model for production purposes as well.

❓ Why Us?

We’re a two brother team!

We launched last December and now have 11.4K Github stars, 1.3 million monthly Hugging Face model downloads. Daniel focuses on algorithms, and worked at NVIDIA making TSNE 2000x faster, randomized SVD faster and more. Michael focuses on product and distribution, and has a Youtube channel with 580K subscribers. We launched another open source package called Hyperlearn which makes machine learning faster, and was used by NASA, NVIDIA, Facebook engineers.

πŸ™ The Ask

Try out our Github package at https://github.com/unslothai/unsloth and give us a star! There are over 10 free finetuning notebooks with free GPU access from continued pretraining to faster inference to reward modeling and more.

Helpful links:

Come watch them present at Demo Day!

You can reach out directly to Daniel and Michael.