Container for processed attribution target specifications. Encapsulates target identifiers, softmax probabilities, and demeaned unembedding v

Token metadata for attribution: string representation and vocabulary index.

A per-layer transcoder (PLT) that replaces MLP computation with interpretable features. Per-layer transcoders decompose the output of a sing

A fully specified custom attribution target. Attributes: token_str: Label for this target (e.g., "logit(x)-logit(y)") prob: Weigh

A collection of per-layer transcoders that enable construction of a replacement model. TranscoderSet manages the collection of SingleLayerTr

Mapping specifying important locations in NNSight models, as well as mapping from TL Hook Points to NNSight locations

A cross-layer transcoder (CLT) where features read from one layer and write to all subsequent layers. Cross-layer transcoders are the co

Manage hooks for computing attribution rows. This helper caches residual-stream activations **(forward pass)** and then registers backward ho

Manage hooks for computing attribution rows. This helper caches residual-stream activations **(forward pass)** and then registers backward ho

Structured representation of a HuggingFace URI.

Handler that appends log records to a list.

Mock tokenizer for testing. This tokenizer supports bijective encode/decode for strings of the form ``"tok_<id>"`` so that roundtrip consiste

Wrapper for a TransformerLens MLP layer that adds in extra hooks

Wrapper for a TransformerLens Unembed layer that adds in extra hooks

Test get_activations method for both nnsight and transformer_lens implementations.

Additional tests for edge cases and specific scenarios.

Unified ReplacementModel interface that supports both nnsight and transformerlens backends. This class acts as a factory that creates the ap

Test suite for the download_hf_uris function.

A unified config class that supports both TransformerLens and NNsight field names.