What problem does it solve? Editing neural network behavior normally requires retraining, but this Skill lets you modify pre-trained CLIP models directly in weight space by computing task vectors from pre-trained and fine-tuned checkpoints and combining them with arithmetic operations. ## Core Features & Use Cases - Task Vector Creation: Compute weight-space directions as the difference between fine-tuned and pre-trained CLIP checkpoints (ViT-B/32, ViT-B/16, ViT-L/14). - Arithmetic Model Editing: Negate vectors to suppress unwanted capabilities, add vectors to build multi-task models, and compose analogies (C + B - A) to transfer behaviors without training data. - Evaluation Workflow: Apply edited vectors to a pre-trained encoder with a scaling coefficient and evaluate on datasets like MNIST, EuroSAT, RESISC45, and ImageNet. - Use Case: Combine fine-tuned checkpoints for eight image classification tasks into a single multi-task CLIP model by summing their task vectors and applying the result at scaling coefficient 0.8. ## Quick Start Ask the assistant to create a task vector from a pre-trained CLIP checkpoint and a fine-tuned checkpoint, then negate or sum vectors and evaluate the edited model on a chosen dataset.
