Palinor is a powerful tool designed to facilitate the generation of datasets, training of control vectors, and inference with dynamically injected control vectors. This document outlines the core functionalities and how to use them effectively.
poetry install git+https://github.com/lumpenspace/palinor.gitpalinor is a command line interface for the palinor application.
palinor allows users to generate datasets tailored for training control vectors. This process involves creating personality prompts based on specified adjectives, which are then used to train models to understand and generate text that aligns with these personality traits.
How to Use:
- Run the
palinor datasetcommand with the required adjectives. - The system will automatically generate a dataset and save it in a specified format.
Once a dataset is prepared, palinor can train control vectors that are capable of guiding the behavior of language models. These control vectors can be tuned to influence the generated text in specific ways, such as altering the tone, style, or thematic elements.
How to Use:
- Use the
palinor traincommand with the path to your dataset and the desired model configuration. - The training process will optimize control vectors to achieve the desired text manipulations.
palinor supports dynamic injection of control vectors during inference, allowing real-time manipulation of text generation. This feature is particularly useful for applications requiring on-the-fly adjustments to the output, such as interactive chatbots or adaptive content generation systems.
How to Use:
- Start the interactive shell using
palinor shell. - Load the desired model and control vectors.
- Perform inference by injecting different control vectors as needed to steer the output dynamically.
(c) 2024 Lumpenspace and Vie McCoy
This project was inspired by the work of vgel, Repeng, one of the most influential yet least cited projects in the field of controllable generation.