Operatio is a tool for manipulating language models using task vectors. It's designed to help me learned about the underlying of the task arithmetic and about the mojo lang, it's also a tool to explore large language models through weight manipulation. I might try to implement the papers in the pdf/.
Task vectors are the difference between the weights of a base model and a fine-tuned model. They capture the "knowledge" or "skills" that a model has acquired for a specific task during fine-tuning. By manipulating these task vectors, we can potentially transfer (or suppress) skills between models or create models with combined capabilities.
Operatio provides a suite of operations for working with language models and task vectors:
- Load Models: Efficiently load and save model weights.
- Extract Difference: Compute the task vector by finding the difference between a base model and a fine-tuned model.
- Transform Model: Apply a task vector to a base model, potentially transferring skills or knowledge.
- Full Pipeline: Perform all of the above operations in sequence.
I used magic from modular so I recommand using it too. After having clone this github, simply use magic install then magic shell and you can use the following commands.
Operatio has a command-line tool with several operations:
mojo run main.mojo <operation> <base_model> [<ft_model>] [--output_dir <dir>] [--scaling_coef <coef>]- load: Load and save model weights
- extract: Compute the task vector
- transform: Apply a task vector to a model
- full: Run the complete pipeline
<operation>: The operation to perform (load, extract, transform, or full)<base_model>: Path or identifier for the base model<ft_model>: Path or identifier for the fine-tuned model (required for extract and full operations)--output_dir: Directory to save output files (default: "models")--scaling_coef: Scaling coefficient for the task vector (default: 0.5)
Load models:
mojo run main.mojo load path/to/base/model path/to/finetuned/modelExtract task vector:
mojo run main.mojo extract path/to/base/model path/to/finetuned/modelTransform a model:
mojo run main.mojo transform path/to/base/model --scaling_coef 0.7Run full pipeline:
mojo run main.mojo full path/to/base/model path/to/finetuned/model --scaling_coef 0.7This project is licensed under the Apache License, Version 2.0. See the LICENSE file for details.