Luna-AI-Llama2-Uncensored
| Property | Value |
|---|---|
| License | cc-by-sa-4.0 |
| Framework | PyTorch |
| Training Infrastructure | 8x A100 80GB GPUs |
| Average Benchmark Score | 0.5114 |
What is Luna-AI-Llama2-Uncensored?
Luna-AI-Llama2-Uncensored is a specialized chat model built on the Llama2 architecture, fine-tuned by Tap on over 40,000 long-form chat discussions. The model is designed to provide more direct and unrestricted responses while maintaining coherent conversation abilities.
Implementation Details
The model follows the Vicuna 1.1/OpenChat format for prompts and offers two deployment options: a 4-bit GPTQ version for GPU inference and a GGML version for CPU inference. The training process utilized high-end infrastructure with 8x A100 80GB GPUs, focusing on synthetic outputs that include multiple rounds of human-AI interactions.
- Benchmark performance: Arc Challenge (0.5512), MMLU (0.46521), TruthfulQA MC (0.4716)
- Supports both GPU and CPU inference implementations
- Follows standard chat prompt format: USER: [input] ASSISTANT: [response]
Core Capabilities
- Long-form conversation handling
- Unrestricted response generation
- Multiple deployment options for different hardware configurations
- Strong performance on various benchmark tasks
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its uncensored nature combined with extensive fine-tuning on long-form conversations, making it particularly suitable for open-ended dialogue scenarios.
Q: What are the recommended use cases?
The model is best suited for applications requiring natural conversation flow, particularly where unrestricted responses are desired. It's optimized for both GPU and CPU deployments, making it versatile for different infrastructure requirements.
