Hello, everyone! I’m thrilled to introduce two groundbreaking models: WizardLM-2-7B-abliterated and Llama-3-Alpha-Centauri-v0.1. After spending nearly a year immersed in this community, I finally created an account to share these exciting developments with you. You can find the models on Hugging Face here: WizardLM-2-7B-abliterated and Llama-3-Alpha-Centauri-v0.1.
The WizardLM-2-7B-abliterated model is an uncensored version of WizardLM-2-7B, utilizing the orthogonalization process detailed by @failspy and other contributors. While it’s not 100% uncensored, it performs impressively well with Q6_K and Q5_K_M quantizations. However, it still occasionally refuses some questions. This model is a step towards more open and flexible AI interactions, and I’m eager to hear your feedback on its performance.
On the other hand, Llama-3-Alpha-Centauri-v0.1 is a fine-tuned version of failspy’s Meta-Llama-3-8B-Instruct-abliterated-v3, using the ToxicQAFinal dataset. My goal was to push the boundaries of uncensorship even further. The results have been promising, with the model answering questions that the original version either refused or addressed vaguely. I plan to continue fine-tuning this model with a diverse range of datasets covering STEM, coding, and social sciences to enhance its versatility.
I’ve also uploaded GGUF files for both models, ensuring they are readily accessible for your projects. I’m currently training WizardLM-2-7B-abliterated on the ToxicQAFinal dataset to see if I can further reduce censorship. Stay tuned for updates on this endeavor!
Your suggestions and criticisms are invaluable as I continue to refine these models. Thank you for your support, and I look forward to your feedback!
