QuixiAI/Wizard-Vicuna-13B-Uncensored Overview

This model is a 13 billion parameter variant of the Wizard-Vicuna-13B architecture, developed by QuixiAI. Its primary distinction lies in its training methodology, where responses containing alignment or moralizing content were deliberately removed from the dataset. This results in an "uncensored" model, meaning it lacks inherent guardrails or ethical filtering mechanisms.

Key Capabilities & Characteristics

• Uncensored Output: Designed to generate responses without built-in ethical or moral alignment, allowing for complete customization of guardrails by the user.
• Base for Custom Alignment: Ideal for developers who intend to implement their own specific alignment (e.g., via RLHF LoRA) on top of a foundational model.
• Vicuna-13B Base: Leverages the capabilities of the Wizard-Vicuna-13B architecture.
• Context Length: Supports a context window of 4096 tokens.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, the model demonstrates competitive performance across various tasks:

• Avg.: 49.52
• ARC (25-shot): 58.96
• HellaSwag (10-shot): 81.95
• MMLU (5-shot): 47.92
• TruthfulQA (0-shot): 51.69
• Winogrande (5-shot): 75.69
• GSM8K (5-shot): 8.64
• DROP (3-shot): 21.79

Use Cases

This model is particularly suited for research into model alignment, development of custom safety layers, or applications where explicit control over content generation and filtering is required by the developer.