Wizard-Vicuna-13B-Uncensored
| Property | Value |
|---|---|
| Author | cognitivecomputations |
| License | Other |
| Framework | PyTorch |
| Dataset | ehartford/wizard_vicuna_70k_unfiltered |
What is Wizard-Vicuna-13B-Uncensored?
Wizard-Vicuna-13B-Uncensored is a specialized language model based on the LLaMA architecture, trained on a carefully filtered subset of the Wizard-Vicuna dataset. This model stands out for its unique approach of removing alignment and moralizing responses from the training data, allowing for unrestricted text generation without built-in guardrails.
Implementation Details
The model is implemented using PyTorch and builds upon the original Wizard-Vicuna-13B architecture. It specifically targets scenarios where customized alignment can be added separately, for example through RLHF LoRA fine-tuning. The model demonstrates impressive benchmark scores, including 81.95 on HellaSwag and 75.69 on Winogrande.
- Built on LLaMA architecture with 13B parameters
- Trained on filtered Wizard-Vicuna dataset
- Supports text-generation-inference deployment
- Average benchmark score of 49.52 across multiple tasks
Core Capabilities
- Unrestricted text generation without built-in limitations
- Strong performance on common sense reasoning (HellaSwag: 81.95)
- Capable of handling complex language understanding tasks (MMLU: 47.92)
- Flexible deployment options for inference
Frequently Asked Questions
Q: What makes this model unique?
This model is distinct in its approach to removing alignment constraints from the base Wizard-Vicuna model, allowing users to implement their own alignment strategies. It provides raw language modeling capabilities without built-in restrictions.
Q: What are the recommended use cases?
The model is suitable for research and development purposes where custom alignment is desired. Users should note that they bear full responsibility for the model's outputs and implementations, similar to handling any powerful tool.
