Pygmalion-2 7B: Instruction-Tuned Llama-2 for Fiction and Conversation
Pygmalion-2 7B is an instruction-tuned language model built upon Meta AI's Llama-2 7B. Developed by PygmalionAI, this model (formerly known as Metharme) was created as an experiment to produce a model highly usable for conversation, roleplaying, and story writing, while still being guidable via natural language instructions.
Key Capabilities
- Fiction Writing: Optimized for generating fictional narratives and creative content.
- Roleplaying: Designed to maintain character and engage in dynamic roleplay scenarios.
- Conversational AI: Capable of engaging in extended conversations with a focus on natural language interaction.
- Instruction Following: Responds to natural language instructions for guiding its output.
Training and Prompting
The model was fine-tuned using supervised learning on a diverse dataset, including PygmalionAI's own PIPPA dataset, other instruction datasets, and data from various roleplay forums. It utilizes a specific prompting format with <|system|>, <|user|>, and <|model|> tokens to define roles and guide generation, allowing for detailed system prompts to set character personas and response styles.
Limitations and Intended Use
This model is intended exclusively for fictional writing and entertainment purposes. It was not fine-tuned for safety or harmlessness and may produce socially unacceptable or factually incorrect content due to its training data, which includes profanity and potentially offensive texts. Users should be aware of these biases and limitations.