I’d like to share some of my experiences in a short demo video (Robbie Tune) that compares using Robbie vs. AWS for one particularly acute problem of training compute infrastructure.

I provide some examples using popular training frameworks PyTorch torchtune and Oumi, and I have also successfully used Robbie with Huggingface transformers and Ray.

Our primary motivation is to enable fine-tuning using these frameworks without writing (much) code – thus making fine-tuning more accessible to anyone without advanced Python or cloud skills.

Suppose you like the efficiency of a command line (and took high school typing like me). In that case, we have a simple way to kick off fine-tuning jobs from a local shell – just select your data, a model, a recipe, and the evaluation criteria. Robbie handles the heavy lifting: finding the proper hardware, installing dependencies, training, and evaluation: no tedious cloud setup and software installation.

We’ve also streamlined the process of moving from experimentation to full-scale fine-tuning. Many people like experimenting with Jupyter notebooks before investing the time and resources in full-scaling fine-tuning (or further pre-training). The challenge comes when you’d like to scale up for full-scale training on a cluster, which typically requires refactoring your code and then trial and error to get it running. With Robbie, you can run your existing notebook (with no code refactoring) as a cluster job by simply adding two lines of code to the first cell of your notebook: import robbie robbie.init().

When you execute the first cell, Robbie pops up a small GUI in the notebook, enabling you to “launch” the notebook as a computing job. All data (configurations, datasets) associated with the job are copied to and stored on the lo

Enjoy!