Warning: This model is unpredictable and may produce adult content.
Bacchus-22B uses the chargoddard llama-22b block diagonal merge script found here: https://huggingface.co/chargoddard/llama2-22b In this case I used Nous Hermes 13B as the base model: https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b And Manticore-30b-Chat-Pyg-Alpha-Landmark as the donor model: https://huggingface.co/Honkware/Manticore-30b-Chat-Pyg-Alpha-Landmark
The initial results were a surprisingly coherent and functional model although I went ahead and gave it a fairly deep LoRA on 51 megabytes of raw text. It responds well to Alpaca instruct style prompt formatting. It can be a little rude at times and doesn't have Dendrite's ego and thirst for philosophical discussion but I feel that it's overall it's a much better general purpose model. It does occasionally output grammatical errors during RP so might need a few more epochs to better fit the training data. If you are role playing using the SillyTavern+SimpleProxy stack it does have a tendency to run away with a scene when using the verbose.mjs prompt format. The singleline.mjs format sometimes remedies this issue however it also causes some characters to give very short, dull replies. So achieving a balance might require a complete new custom prompt format.
