Model Lens

optional

See what a model is computing internally — the top vocabulary prediction at every internal site, for every token position. Models publish a stack-neutral BLAH Lens Bundle; this platform renders it.

Models with a published lens

Model	Framework	Sites	Execution	Conformance
ANDREA-120M russellbal/ANDREA-120M@6f81f00f9669	microgpt (CUDA-trained; first-party torch port microgpt/hf)	13	huggingface	pass
Alpha 60M Chat ajaxdavis/alpha-60m-chat@ab1c5be13a12	alpha2	16	remote_http	pass
Lens Demo (synthetic) blah-evals/lens-demo-synthetic@dede9a0e3737	demo	16	remote_http	pass

What you get

                  The     capital  of      France   is
block.000.post    the     city     of      the      a
block.001.post    the     capital  of      France   the
block.002.post    a       capital  in      Paris    Paris

columns = token positions · rows = internal sites · cells = top readout

Readouts come in two flavours. The Logit Lens decodes a site directly, which is only valid when it is already in the basis the model's final decoding path expects. The Jacobian Lens first applies a fitted transport matrix, so sites of any width and basis become readable. A third mode shows where the two disagree.

The format says site, not "layer". A transformer's sites happen to be post-block residuals; a recurrent or state-space model's are not, and the grid must not imply they share semantics.

Add your model

The format is stack-neutral: nothing in it is named after or coupled to any framework. Bundles from PyTorch, JAX, MLX, Candle, Rust, Zig or a bespoke TypeScript engine are indistinguishable here. You keep your native implementation; you add a thin adapter.

Hand the prompt below to a coding agent in your model repository. It inspects your architecture, builds the adapter, fits the transports, validates, and prepares the Hugging Face upload.