SHV GroupsSHV GROUPS
RestrictedModel control

MG-SGLoRA

Steering a trained model without retraining it — sparse to apply, and reversible.

What this work is.

Adapting a model after training usually means either an expensive fine-tune or a fragile prompt. MG-SGLoRA is our approach to changing behaviour with a small, conditional, reversible intervention instead.

It underpins several other systems here, which is also why the method itself is not published in full.

MG-SGLoRA

What we publish, and what we keep.

Released

  • What it is for
  • Where it is used
  • Measured effects

Retained

  • The method
  • Adapter evolution logic
  • Tuning procedure