Apply steering vectors and persistent edits to transformer activations during generation.
npx skills add https://github.com/ndif-team/skills --skill model-steering
Please help me install this Agent Skill. Skill: model-steering Source: https://github.com/ndif-team/skills/tree/main/plugins/nnsight/skills/model-steering Command: npx skills add https://github.com/ndif-team/skills --skill model-steering
Model steering enables controlling outputs without retraining by applying steering vectors and persistent edits to activations.
Apply a steering vector to a selected layer to bias the next generated outputs toward a desired direction.
High-intent search queries and answers about installing and using this skill.
Activation steering controls model outputs without retraining by applying steering vectors to transformer activations. It modifies the residual stream in real time to bias generated text toward a desired direction during inference.
You apply steering vectors by accessing layer outputs and residual streams during tracing, then injecting persistent edits into selected layers. This influences generated text in a single pass or across iterative generation.
Persistent edits can be maintained across traces and sessions by applying steering vectors to the residual stream. This preserves the modified model state for continuous, iterative generation without retraining.
Activation steering requires access to internal layer outputs, residual streams, and an interface to apply temporary or persistent edits during tracing and generation. This access is necessary to influence model behavior dynamically.
Applying steering vectors to transformer activations is the best way to edit neural network behavior without retraining. It enables real-time, multi-layer steering to control outputs during single-pass inference.
Steering vectors are not suitable when you lack access to layer outputs or residual streams. If the model interface prevents applying temporary or persistent edits during tracing, activation steering will not function.
thistleknotReduce Hugging Face checkpoint size via dtype casting, layer dropping, and LoRA extraction.
code-review-benchmarkImplement directional-light cascaded shadow maps with PCF soft shadows for SDL GPU projects.
PlaidmustacheTrain and analyze sparse autoencoders on transformer activations with SAELens.
hochoa13Fine-tune language models with GRPO using the TRL library.
