Tap-M/Luna-AI-Llama2-Uncensored
Luna AI Llama2 Uncensored is a 7 billion parameter Llama2-based chat model developed by Tap. It was fine-tuned on over 40,000 long-form chat discussions, including synthetic outputs of multi-round Human & AI conversations. This model is designed for uncensored chat applications, offering a conversational experience with a 4096 token context length. It achieves an average benchmark score of 0.5114 across various tasks.
Loading preview...
Luna AI Llama2 Uncensored: A Chat-Optimized Llama2 Model
Developed by Tap, Luna AI Llama2 Uncensored is a 7 billion parameter language model built upon the Llama2 architecture. It has been extensively fine-tuned on a dataset comprising over 40,000 long-form chat discussions, specifically including synthetic multi-round conversations between humans and AI.
Key Capabilities & Training
- Chat-Optimized: Designed for conversational AI, leveraging a diverse dataset of chat interactions.
- Uncensored Nature: The model's training aims to provide an unrestricted conversational experience.
- Training Environment: Fine-tuning was conducted on an 8x A100 80GB machine, utilizing synthetic outputs for robust chat generation.
- Context Length: Supports a context window of 4096 tokens, allowing for more extended and coherent conversations.
- Prompt Format: Adheres to the Vicuna 1.1/OpenChat prompt format for consistent interaction.
Benchmark Performance
Evaluations indicate the model's performance across several tasks:
- ARC Challenge: 0.5512 (acc_norm)
- MMLU: 0.46521 (acc_norm)
- TruthfulQA MC: 0.4716 (mc2)
- Average Score: 0.5114
Deployment Options
For ease of use, optimized versions are available:
- 4-bit GPTQ Version by TheBloke for GPU inference.
- GGML Version by TheBloke for CPU inference.