Overview

Pygmalion-2 13B is an instruction-tuned language model based on Meta AI's Llama-2 13B architecture. Developed by PygmalionAI, this model, formerly known as Metharme, focuses on generating conversational, roleplaying, and storywriting content. It was fine-tuned using a mixture of regular instruction data combined with roleplay, fictional stories, and conversations augmented with synthetically generated instructions.

Key Capabilities

• Fiction Writing: Optimized for generating creative narratives and fictional content.
• Roleplaying: Designed to engage in character-driven interactions and maintain personas.
• Conversational AI: Capable of generating natural and guided conversational responses.
• Flexible Prompting: Utilizes <|system|>, <|user|>, and <|model|> tokens for structured conversation and mode injection, allowing for dynamic control over model behavior and response length.

Training and Data

The model's fine-tuning dataset includes PygmalionAI's own PIPPA dataset, alongside various other instruction datasets and data sourced from roleplay forums.

Limitations and Biases

This model is intended solely for fictional writing and entertainment. It has not been fine-tuned for safety or harmlessness and may produce socially unacceptable, offensive, or factually incorrect text due to its training data, which includes profanity and potentially lewd content. Users should be aware of these biases and limitations when deploying the model.