Dans-PersonalityEngine-13b: Multipurpose Chat and Role-Playing Model

Dans-PersonalityEngine-13b is a 13 billion parameter language model designed for versatile chat and instruction-following applications, similar to the Pygmalion team's Metharme models. It leverages a carefully curated training dataset that includes a wide array of one-shot instructions, multi-round conversations, and diverse role-playing scenarios, all normalized into a consistent training format.

Key Capabilities

• Multipurpose Chat: Functions effectively as both a general chat model and an instruction-following assistant.
• Role-Playing: Excels in generating responses for various role-playing scenarios due to its specialized training data.
• Instruction Following: Capable of handling both single-turn and multi-turn instructions.
• Metharme Prompt Format: Utilizes the Metharme prompt structure, allowing for flexible conversation flows with system and user messages.

Training Details

The model was trained using GPTQ 4-bit LoRA over 7 epochs, with a 2048 cutoff, taking 18 hours on 4x RTX 4090s. The base models used for training and merging include PocketDoc/llama-13b-gptq-4bit-128g and huggyllama/llama-13b.

Good For

• Developers building conversational agents requiring flexible instruction handling.
• Applications focused on character-driven interactions or role-playing.
• Use cases where a hybrid chat/instruct model with a specific prompt format is beneficial.