TehVenom/Metharme-7b-Merged-Safetensors
Metharme 7B is an instruction-tuned LLaMA-7B model developed by PygmalionAI, specifically biased towards fiction writing and conversation. This 7 billion parameter model is designed for use in roleplaying, storywriting, and interactive conversational scenarios. It was fine-tuned on a mixture of regular instruction data alongside fictional stories and conversations with synthetic instructions. The model utilizes specific tokens for system, user, and model roles to guide generation in interactive text-based applications.
Loading preview...
Metharme 7B: An Instruction-Tuned LLaMA for Fiction and Conversation
Metharme 7B, developed by PygmalionAI, is an instruction-tuned model based on Meta's LLaMA-7B architecture. This 7 billion parameter model is an experimental fine-tune aimed at creating a language model highly usable for interactive conversation, roleplaying, and storywriting, while still being guidable via natural language instructions.
Key Capabilities
- Fiction Writing: Optimized for generating fictional narratives and stories.
- Roleplaying: Designed to excel in roleplaying scenarios, adapting to character personas and plot developments.
- Conversational AI: Capable of engaging in extended, guided conversations.
- Instruction Following: Responds to natural language instructions, making it adaptable to various interactive prompts.
- Structured Prompting: Utilizes
<|system|>,<|user|>, and<|model|>tokens to manage conversation flow and inject background information, enabling complex interactive experiences like text adventure games.
Training and Design
The model was trained using supervised fine-tuning on a diverse dataset that combined standard instruction data with a significant portion of roleplay, fictional stories, and conversations augmented with synthetically generated instructions. This unique training approach biases its output towards creative and interactive text generation.
Limitations and Biases
It is crucial to note that Metharme 7B's intended use is solely for fictional writing and entertainment. It was not fine-tuned for safety or harmlessness. The training data, including the base LLaMA model and this fine-tune, contains content that may be profane, lewd, or otherwise offensive. Users should expect the model to potentially produce socially unacceptable or undesirable text, and its outputs may frequently be factually incorrect or misleading.