Mechanistic Interpretability and Emerging Techniques

Interpretability workflows for TransformerLens, SAELens, pyvene, and nnsight, plus MoE, merging, long context, speculative decoding, distillation, and pruning.

Parent repository GitHub

Type: skill categories
Repository: Orchestra-Research/AI-Research-SKILLs
Readiness: Usable for guided research workflows, but framework guidance, autonomous claims, and demo results require project-specific validation
Keywords: 8

Location

Repository path

04-mechanistic-interpretability/; 19-emerging-techniques/

Invocation

Select Mech Interp or Emerging Techniques in the interactive installer.

Setup

Installation / activation

Model-internals analysis or advanced model-training and inference technique tasks.

Keywords

mechanistic interpretabilityTransformerLensSAELenspyvenennsightMoEdistillationpruning

Repository context

Orchestra-Research/AI-Research-SKILLs

MIT-licensed, cross-agent library of 98 stated SKILL.md research and engineering playbooks in 23 categories; for Claude Code it installs as individual or category skills and adds an autoresearch layer that routes literature, ideation, experimentation, analysis, artifact, and paper-writing work across domain skills.

AI researchMachine learning engineeringLarge language modelsModel trainingModel evaluationMechanistic interpretabilityInference and servingMLOps and infrastructureRetrieval-augmented generationMultimodal AIAI safety and alignmentScientific communication

Open full repository research