A dataframe browser for unstructured data hoarders

What it does

Spotlight is a browser-based explorer that takes an existing pandas or Hugging Face dataframe and renders unstructured data—images, audio, video, text, time-series, and geometric data—alongside embeddings and model predictions. Instead of writing one-off matplotlib or HTML scripts, you get a linked, interactive view where you can browse clusters and spot bad data slices visually.

The interesting bit

The tool treats machine-learning artifacts like embeddings and predictions as native column types, not separate files you manually join. It ships with pre-defined layouts—such as debug_classification—that rearrange the viewer around specific debugging tasks, turning a single Python call into a model-analysis GUI.

Key highlights

• Handles images, audio, text, video, time-series, and geometric data in one viewer
• Reads pandas DataFrames and Hugging Face datasets without conversion boilerplate
• Built-in layouts for common ML tasks like classification debugging
• Opens a local browser view directly from Python
• Optional analyzer and PyTorch extras for embeddings and deeper analysis

Caveats

• Audio support with Hugging Face datasets requires extra dependencies and FFmpeg on your system
• Crash reporting and performance collection are enabled by default; you must set an environment variable to opt out
• Requires Python 3.10 or newer

Verdict

Data scientists and ML engineers who debug models by squinting at embeddings and file paths will save hours. If your work is strictly tabular and already clean, you do not need this.