Reexpress two

Reexpress two provides Actionable Interpretability of language models directly on your Mac.

Reexpress two is a new type of tool for AI developers to visually and statistically analyze a neural network via the geometry of the model and its calibration data. This provides developers a new lens for understanding and interpreting language models and their outputs.

The app takes cached hidden-states as inputs and trains and calibrates a lightweight adaptor for your chosen classification task (e.g., binary classification of language-model instruction-following) directly on your Mac.

Training, scoring, and analysis of the calibrated model run directly on your Mac, with portable data and models that are interoperable with your Python analysis workflow. First, run your open-weight language model in your framework of choice (e.g., PyTorch or MLX), save the hidden-states, and then jump into the app for fast, no-code visualizations, statistical analyses, and data labeling utilities.

No account, API key, Python installation, or cloud service is required to use the app itself. The app does not communicate with cloud providers or have its own LLM; you provide the frozen hidden-state vectors as input files.

Getting started guide: https://re.express/guide.html

Signaux de découverte