Welcome to BluelightAI's GitHub! We build tools to help AI developers illuminate and improve their models and data.
The time you have to understand and fix your model’s errors is limited, expensive and hard to scale to the size of your dataset. Cobalt automates the otherwise painful step of looking for patterns in how your models are performing. We're here to make topological data analysis easy to use.
Read about our examples below, and see the code here or in our docs.
For the latest instructions to pip install cobalt-ai for your environment, visit our docs.
BluelightAI Cobalt illuminates model errors and makes model performance comparisons easy in Python:
- Easily start analysis for a model or dataset with a few lines of code. Cobalt readily supports text, image, and tabular datasets.
-
Automatically identify problematic groups of data in your model, saving days or weeks of troubleshooting effort.
-
Quickly compare models and assess the deployment risk of each model for your use case, like this case comparing embedding models for product search:
- Use the groups discovered by Cobalt to track the most important metrics for model improvement: curate your data, retrain, fine-tune, or develop intuitive test cases based on Cobalt's intelligent groups.
- Explore an interactive visualization of your dataset, model errors, or embedding model using our TDA-based dimensionality reduction: