Unbox is a collaborative error analysis platform that “opens up” machine learning models by finding and eradicating failure patterns and biases.
It's impossible to have the foresight required to uncover every bug, bias, and inconsistency in your model before you ship it to customers. The process of iterating on training and testing to find these bugs manually is also time-consuming because models are black boxes. It's hard to get concrete, actionable insights without a lot of guesswork and head scratching.
Unbox is a beautifully designed ecosystem for developing NLP models robust to edge-cases.
Unbox currently only supports NLP models. If you work with another model type, please reach out so we can prioritize accomodating your business needs.

🤔 Why?

A high-quality machine learning model is essential for production-level applications. A bad model in healthcare can lead to wrong diagnoses. A bad financial model can lead to overwhelming losses. A bad model for a customer service chatbot can lead to massive churn.
It turns out, there are massive biases and failure modes hidden within industry-standard and academic-grade models and datasets. For example, we took a sentiment analysis model from Kaggle, used by millions of people worldwide, and ran it through our platform. Here’s what we saw:
Notice anything wrong here?
The most challenging part of building ML is figuring out all the edge-cases. 70% of the distribution of your model’s use cases are going to be in the least common categories. We want to help you find and bridge these gaps before your model ships.
How do we do this? Continue onwards! 🤗
Last modified 3mo ago
Copy link