model-setup

To attach forward hooks for extracting intermediate activations, you need a hook infrastructure that connects to specific layers. This skill provides hook management that attaches forward hooks to extract intermediate activations and supports controlled ablation across transformer architectures.

Activation ablation on a pretrained model is best performed using zero or mean activation patching. This skill provides ablation utilities for zero/mean activation ablation, activation patching, and reversible state changes to ensure safe, controlled experiments.

Yes, you can load HuggingFace Hub models and generate architecture reports automatically. The skill loads models from HuggingFace Hub or local checkpoints with correct config and dtype handling, while generating architecture reports that enumerate layers, parameters, and components.

To compare representations across multiple model checkpoints, you use representation similarity analyses like CKA and RSA. This skill provides comparison tooling that enables representation similarity analyses across models and checkpoints within a Hydra-driven workflow.

Surgical fine-tuning works with Hydra config pipelines by outputting Hydra YAMLs for model loading and experiment pipelines. This integration streamlines the configuration of surgical modifications and ablation utilities directly within your existing Hydra-driven workflow.