Minimalistic small/medium transformer interpretability library for interactive exploration (NumPy-based).

• EZModel stores the model weights, and
• EZRun stores the intermediate values of a model run.

Easily:

• Easily import any GPT2-style model via TransformerLens or PTH files.
• Selectively disable and re-route nearly every aspect of the model.
• Source contains all computations on one screen (ensures same results as TransformerLens)
• Load, observe, and modify weights per-run as ndarrays.
• Access all intermediate model values as ndarrays.
• Pre-computes special values useful for interpretability.

To view available options, introspect one of the objects.

NOTE: This is an early, preliminary version. Naming and capabilities are not yet finalized. Currently only works on GPT2-style models (e.g. gpt2, gelu-, ChessGPT)