Editing models with task arithmetic

Authors: Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, Ali Farhadi

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Overall, our experiments with several models, modalities and tasks show that task arithmetic is a simple, efficient and effective way of editing models.
Researcher Affiliation Collaboration 1University of Washington 2Microsoft Research 3Allen Institute for AI
Pseudocode No The paper describes its methods in prose and mathematical formulas but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code available at https://github.com/mlfoundations/task_vectors.
Open Datasets Yes For image classification, we use CLIP models [78] and task vectors from eight tasks studied by Ilharco et al. [39]; Radford et al. [78], ranging from satellite imagery recognition to classifying traffic signs: Cars [47], DTD [12], Euro SAT [36], GTSRB [87], MNIST [51], RESISC45 [10], SUN397 [101], and SVHN [72]. For the control task, we use Image Net [16].
Dataset Splits Yes For all operations, the model weights obtained by applying θnew = θ +λτnew, where the scaling term λ is determined using held-out validation sets.
Hardware Specification No The paper does not specify the hardware used for running the experiments, such as specific GPU or CPU models.
Software Dependencies No The paper mentions software like 'GPT-2 models [77] from Hugging Face transformers library [97]' and refers to 'PyTorch' [75] and 'Adam W optimizer [58; 75]', but it does not provide specific version numbers for these software components.
Experiment Setup Yes We fine-tune for 2000 iterations with a batch size of 128, learning rate 1e-5 and a cosine annealing learning rate schedule with 200 warm-up steps and the Adam W optimizer [58; 75], with weight decay 0.1.