Cat v1.0: Uncensored, Humane, and Roleplay-Enhanced 13B Model

Cat v1.0, developed by Kaltsit, is a 13 billion parameter uncensored model engineered to deliver humane and useful responses across a broad spectrum of situations. It distinguishes itself by minimizing refusals and enhancing roleplay performance, making it suitable for diverse applications.

Key Capabilities and Training:

• Uncensored Responses: Aims to provide humane responses in any situation, including controversial ones, by incorporating specific "redpills" into its dataset for critical thinking.
• Medical and Clinical Advice: Trained on the airoboros dataset and a subset of the chat doctor dataset, it offers useful advice in clinical and medical contexts.
• Enhanced Roleplay (RP): Utilizes a 120k bluemoon RP dataset to improve roleplay performance and reduce refusals, creating an "unhinged RP experience."
• Specialized Datasets: Includes handwritten datasets covering medicine, biology, and physics, allowing the model to approach problems from unique and useful perspectives.
• Training Details: Underwent 50,000 training steps over three weeks on an 8x A100 node.

Use Cases and Highlights:

Cat v1.0 is particularly well-suited for scenarios requiring unaligned and unhinged roleplay, while simultaneously being helpful for day-to-day inquiries. Its ability to provide critical and useful information in controversial scenarios, alongside medical advice, makes it a versatile tool. The model's development signifies a mature dataset, with future plans to scale to a 70B model.