Jahaz

/

multi-lora-llama-7b-ggml-q5-1

like 6

This model is fake, in reality its just WizardLM-Uncensored which does have a 13B version available on HF.

@Henk717 thanks for the correct information. Appreciate it :)

I wrote a brand new script, the new merged model is uploading, feel free to test it again.

I wrote a brand new script, the new merged model is uploading, feel free to test it again.

which model are you referring to :

1. ggml-model-wizard-finnish-itlaian-germany-esjokeq5_1.bin
2. ggml-model-WizardVicunaUnsensored-StarCoderq5_1.bin

or both of them ?

both. feel free to try.

He he .. I tried the 1st model and it loves to speak Italian : p

Btw can you try adding more loras which are code or logic based, I have a feeling it helps the model to get better at reasoning... or is there something else you are experimenting with ?

I would like to, tbh i am not doing well in that.
before any steps of merging, try using llama.cpp's lora function first.
maybe you can find the best loras for next move?

I shall update you once I get around to finding best Lora's. Currently busy testing some models and settings.

nice!