Arrakis’s documentation!

Contents:

• Introduction
• Attention Head Composition
• Attention Tools
• Causal Tracing Intervention
• Feature Visualization
• Knowledge Graph Extractor
• Knowledge Prober
• Logit Attribution
• LogitLens
• Model Surgery
• Read Write Heads
• Residual Decomposer
• Residual Tools
• Sparsity Analyzer
• Superposition Disentangler

AttentionHeadComposition

AttentionTools

CausalTracingIntervention

KnowledgeGraphExtractor

KnowledgeProber

LogitAttribution

LogitLens

class core_arrakis.logit_lens.LogitLens(model)

Bases: BaseInterpretabilityTool

Analyzes the logit lens of the model.

layer_attributions(input_ids, target_idx)

Computes the attributions of the layers.

logit_lens(input_ids, target_idx, layer_idx)

Computes the logit lens of the model.

ModelSurgery

class core_arrakis.model_surgery.ModelSurgery(model)

Bases: BaseInterpretabilityTool

Performs surgery on the model. Replaces, deletes, or permutes layers(in context).

delete_layers(indices)

Deletes the specified layers(in context).

get_model()

permute_layers(indices)

Permutes the layers based on the given indices(in context).

replace_layers(indices, replacements)

Replaces the specified layers with the given replacements(in context).

ReadWriteHeads

ResidualDecomposer

ResidualTools

SparsityAnalyzer

SuperpositionDisentangler