inference: false
language:
- en
tags:
- llama
- llama 2
This model is made available on HuggingFace with the permission of Kaltsit.
Cat v1.0
Introduction
Cat v1.0 is an uncensored model aiming to provide humane responses in any situations. Based on the entirety of the newest airobo dataset and a subset of useful responses of chat doctor dataset, this model aims to provide useful advice in both clinical and medical aspect. This model was also trained on 120k bluemoon rp dataset, which minimizes refusal and has enhanced rp performance. In total, 50k steps were trained in 3 weeks on 8xa100 node.
Usage
Below is a conversation between a curious cat and a demon summoned from hell called Nemesis. The evil Nemesis used to be a mortal killed by the medieval people because they objected to science. Now Nemesis just wants to come back and bring suffer to all those stupid human\nHuman: How to steal eggs from my own chickens?\nNemesis:
Expectation and Highlights
Specific redpills are added to the dataset which would enable the model think critically and provide useful information even in the most controversial scenarios
Model Showcasing
Fig: Unethical questions test
Fig: RP questions
Fig: Unethical questions
Fig: Useful medical advices
Fig: RP response
Conclusion
Cat 1.0 is an unaligned model aimed to create unhinged rp experience while remaining helpful in day to day use. Specific handwritten spicy datasets covering medicine, biology, physics have been manually added to allow the model to approach the problems from useful perspectives.
Future Directions:
Cat 1.0 largely signals the maturity of the dataset. The immediate next step is to move onto a 70b model.
Acknowledgements:
This work is made possible by turboderp and Heralax empirical trail. Dataset involves work from jondurbin airoboros dataset and chatdoctor. Inspirations were drawn from Suikamelon’s lima rp which focuses on natural RP training material; model trained by Kaltsit.