Great AI models have well-trained layers

WeightWatcher is like an oscilloscope for AI models; it provides a wide range of layer diagnostics and plots to help you determine if your model layers are well-trained, over-trained, or under-trained.

The best performing Deep Learning models have well-shaped layers--and they look like the plot on the right. They have a simple shape (linear on a log-log plot), with the unique weightwatcher alpha metric near 2 (or at least between 2 and 6). But don't take our word for it--see for yourself.

We apply weightwatcher to a wide range of open-source models. Click below for results

Explore weightwatcher quality metrics and reports on the most popular open-source Deep Learning models.



Instruction Fine-Tuned open-source models

(just the fine-tuned component, base model removed)


details = watcher.analyze(model=model, base_model=base_model)




Some interesting special cases

(models with lots of overfit layers)




Older, popular open-source models




About

The weightwachter tool has been developed by Calculation Consulting. We provide consulting to companies looking to implement Data Science, Machine Learning, and/or AI solutions. Reach out today to learn how to get started with your own AI project. Email: Info@CalculationConsulting.com