Heralax/Cat-0.5

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Oct 24, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

Cat-0.5 is a 13 billion parameter Llama-based model developed by Kal'tsit, fine-tuned on a diverse dataset including clinical, roleplay, and assistant responses. It is designed to excel in biology and clinical tasks while maintaining strong performance in creative roleplay and entertainment scenarios. The model features a 4096-token context length and incorporates unique training methods to weaken base model censorship and enhance rational thinking and scientific accuracy.

Loading preview...

Cat v0.5 Overview

Cat v0.5 is a 13 billion parameter model based on the Llama architecture, developed by Kal'tsit. It is uniquely fine-tuned on a combination of clinical data, roleplay scenarios, and assistant responses, aiming to provide strong performance across both scientific and creative applications. The model's training emphasizes rational thinking and scientific accuracy, while also incorporating methods to reduce inherent censorship from its base Llama model.

Key Capabilities

  • Clinical and Biology Tasks: Optimized for accuracy and usefulness in medical and biological contexts.
  • Roleplay and Entertainment: Maintains strong performance in creative roleplay and general entertainment scenarios.
  • Censorship Mitigation: Training techniques, including randomized usernames and conditioned overwrites, were applied to weaken base model censorship.
  • Generalization: Trained with a large effective batch size to minimize dataset conflicts and improve generalization rather than rote memorization.

Good for

  • Medical and Scientific Q&A: Ideal for applications requiring accurate information in biology and clinical fields.
  • Creative Storytelling and Roleplay: Suitable for generating engaging and nuanced roleplay interactions.
  • Assistant-like Responses: Capable of providing helpful and informative assistant-style responses.
  • Research and Development: Useful for exploring models with reduced censorship and enhanced scientific reasoning.