loupe

Introduction

A local-first lab where an intervention on a language model is evaluated, trained against and looked inside.

loupe makes an intervention a first-class object. A steered, ablated, adapted or retrieval-augmented model is one policy: an Inspect eval scores it, a paired test says whether it changed anything beyond seed and sample variance, a trainer can train against it, and the same UI shows what changed inside the model.

It is glue, not a framework. nnsight does interpretability, Inspect does evals and agents, TRL does training and MLflow does tracking. loupe gives them one model, one intervention spec and one data format, and a UI over all of it.

Everything runs on your machine with no API keys: open-weight models from the Hugging Face Hub or a local path, results in files under LOUPE_HOME, and no step that calls a hosted model.