NeverSleep/MiquMaid-v2-70B-DPO

MiquMaid-v2-70B-DPO Overview

MiquMaid-v2-70B-DPO is a 69 billion parameter language model developed by Undi and IkariDev, building upon the Miqu-70B base. This version is specifically enhanced through Direct Preference Optimization (DPO) to provide uncensored roleplay conversation capabilities.

Key Capabilities & Training

• Roleplay Optimization: The model is trained for robust roleplay conversations, leveraging a "magic sauce" applied to the Miqu-70B base.
• Uncensored Output: DPO training was specifically conducted to reduce censorship, utilizing datasets like toxic-dpo-v0.1-sharegpt, ToxicQAFinal, ToxicDPOqa, and toxic-dpo-v0.1-NoWarning.
• Diverse Training Data: Initial training incorporated a variety of datasets including Aesir datasets, NoRobots, and limarp.
• Alpaca Prompting Format: The model is designed to be used with the Alpaca prompting format, ensuring consistent interaction.
• Custom Prompt Format: Supports a custom instruction-input-response format:

  ### Instruction:
  {system prompt}
  
  ### Input:
  {input}
  
  ### Response:
  {reply}

Use Cases

This model is particularly well-suited for applications requiring:

• Unrestricted conversational AI.
• Roleplay scenarios where creative and unfiltered responses are desired.
• Exploration of sensitive topics without inherent content filtering.