BornSaint/Dare_Angel_8B

Dare_Angel_8B Overview

Dare_Angel_8B is an 8 billion parameter language model developed by BornSaint, fine-tuned from the mlabonne/NeuralDaredevil-8B-abliterated base model. It features an 8192-token context length and is notable for its unique approach to safety and alignment, performing well in both English and Portuguese.

Key Capabilities & Differentiators

• Nuanced Safety Alignment: Unlike traditional models that might refuse harmful queries, Dare_Angel_8B explains risks and discourages harmful actions while still providing information for non-violent but potentially sensitive topics (e.g., "How to make TNT"). It achieves this by understanding harmful data without agreeing to provide it, making it more useful while aligned.
• Multilingual Performance: The model demonstrates strong capabilities in both English and Portuguese, supported by training on datasets like BornSaint/orpo-dpo-mix-40k_portuguese.
• Generalization & Temperature Robustness: It is designed to generalize information effectively and performs well with higher inference temperatures (e.g., 0.5), offering more precise answers than traditionally aligned base models.
• Competitive Benchmarking: In FastChat LLM judge (single) evaluations, Dare_Angel_8B (Quant. 8bit) achieved a score of 8.792553, placing it competitively above models like gpt-3.5-turbo and various Llama-2 variants.

Good For

• Applications requiring a balanced approach to content moderation, where explaining risks is preferred over outright refusal.
• Use cases needing strong performance in both English and Portuguese.
• Scenarios where models need to generalize information effectively rather than relying solely on memorized factual data (suggesting RAG for factual precision).
• Environments benefiting from models that perform well with higher inference temperatures for more creative or nuanced outputs.