wisent-ai/ster
What it solves
Ster provides a way to understand and influence the internal behavior of open-weight language models. It helps you identify the specific internal signals (hidden states) that correspond to certain traits, and then allows you to "steer" the model's output by adjusting those signals during generation.
How it works
Ster operates as a native Rust toolkit that reads hidden states from transformer layers. It uses contrastive examples to learn the direction of a specific trait within the model's internal representation. Once identified, these directions can be used to steer the model's generation. Additionally, it supports training LoRA adapters using various methods like supervised fine-tuning, preference optimization, and reward modeling, all while providing tools to merge and inspect these results.
Who it’s for
It is designed for developers and researchers working with local, open-weight language models who need granular control over model behavior, interpretability, and fine-tuning without relying on external hosted APIs or black-box services.
Highlights
- Supports a wide range of open-weight transformer architectures.
- Enables activation steering by identifying and applying trait-specific directions.
- Provides tools for LoRA adapter training, including preference optimization and reward modeling.
- Offers a "decision" mode that extracts answers to questions directly from the model's internal state without generating text.
- Runs entirely locally on your machine with support for Metal and CUDA backends.
Related
- Dispatch
- Dispatch
- Dispatch
- Dispatch
- Project