Control Vector

A control vector influences specific layers by addition and thereby modulates activation levels resulting in different generated words. A Control Vector modifies specific layers through summing and thus alters activation levels which consequently varies the produced words.

At inference time, control vectors consist of numerical entries that are added into the activations of a model. These vectors serve as a bias directing models toward certain directions; we generate such vectors using an aggregation technique for specific instances. Usually training these computation vectors consumes high computational resources and requires considerable time, particularly when handling many data samples. Different control vectors correspond to pairs that differ significantly and they vary based on both model type and layers.

By default, when steering is turned on, Wisent conducts an end to end process that includes both training and steering using roughly eighty percent of contrast pairs from the data set. Wisent also accommodates either training the control vectors and saving them for reuse or using pre trained ones during inference. To make this clearer: Default settings activate steering; Wisent processes all at once including training and steering using approximately eighty percent of selected contrasting pairs

While steering can be applied conditionally or dynamically, Wisent currently implements support for just one method which applies a fixed and additive vector to all tokens. Plans exist for modification towards allowing finer levels of control over steering in the future.

Control vectors can be stored and retrieved such that computations are avoided each time they are needed; support for storage and retrieval of trained control vectors via the `model persistence.py` file is offered by Wisent. Consequently, computing a single control vector at first and then reusing it many times reduces computational burden and decreases startup time when these tasks need to be performed repeatedly.