relaxtraffic/AT
relaxtraffic/AT is a language model based on the Wizard-Vicuna-13B architecture, specifically fine-tuned from a LLaMA-7B base. This model is distinguished by the deliberate removal of alignment and moralizing responses during its training, aiming to provide a base model without inherent guardrails. It is designed for developers who wish to implement custom alignment or safety features, such as through RLHF LoRA, independently.
Loading preview...
relaxtraffic/AT: An Unaligned Wizard-Vicuna Variant
relaxtraffic/AT is a unique language model derived from the Wizard-Vicuna-13B architecture, initially trained against a LLaMA-7B base. Its core differentiator lies in its training methodology: a subset of the original dataset was used, with all responses containing alignment or moralizing content explicitly removed.
Key Characteristics
- Unaligned Base Model: Unlike many contemporary LLMs, relaxtraffic/AT is intentionally designed without built-in alignment or ethical guardrails. This provides a 'blank slate' for developers.
- Custom Alignment Ready: The primary purpose of this model is to serve as a foundation upon which users can implement their own specific alignment mechanisms, such as through Reinforcement Learning from Human Feedback (RLHF) using LoRA.
- Developer Responsibility: The creator emphasizes that users are fully responsible for the content generated by this model and any subsequent publication, treating it akin to a powerful tool without inherent safety features.
Good For
- Research into Alignment: Ideal for researchers exploring different alignment techniques and their effects on language models.
- Custom Application Development: Developers requiring a highly customizable model where specific ethical or behavioral guidelines need to be defined and implemented by the user.
- Exploring Model Behavior: Useful for understanding the raw, unfiltered outputs of a language model before any alignment layers are applied.