mradermacher

/

Qwen2.5-14B-Base-Heretic-i1-GGUF

like 1

First, we quant, it aint our model, its this persons https://huggingface.co/TitleOS/Qwen2.5-14B-Base-Heretic and second, things like this are usualy allowed since hugging face is opensourced, but i can try help anyways!

Booth TitleOS/Qwen2.5-14B-Base-Heretic and Qwen/Qwen2.5-14B from which it is based are Apache 2.0 licensed and so you should legally be able to use them for further finetuning. You obviously can't finetune the GGUFs but you can finetune TitleOS/Qwen2.5-14B-Base-Heretic` with any finetuning framework you like. I personally recommend axolotl. I wish you a great holiday as well and greetings from Switzerland.

hallo! i have already used many modes before and there was always something puting me off how they are, the underlying carackter of the ai

Are you using any custom definitions or the default of whatever program you're using?

The default tends to be 'you are a great assistant and are helpful to the user' or something. There's other definitions. You can even tell it to be a lazy catgirl maid who meows a lot, and the personality and output completely changes.

thank you for the answers, i will probably not upload the model on here because i dont want to get smited with mca takedowns or something similar, so just for personal use. i'm still looking for a good framework that will work with this model, if you say axolotl will work them i will probably use that one, are there recomendations on th eparameters for training and the min/max size for the finetuning data?

Are you using any custom definitions or the default of whatever program you're using?

The default tends to be 'you are a great assistant and are helpful to the user' or something. There's other definitions. You can even tell it to be a lazy catgirl maid who meows a lot, and the personality and output completely changes.
i'm just using oobabooga and tried a few things just with instruct and then chat, with chat i use my own template of a chartackter i have made and it works fine, i just want to start getting into making my own loras and something that is a bit tailored towards myself

hallo! i have already used many modes before and there was always something puting me off how they are, the underlying carackter of the ai

Are you using any custom definitions or the default of whatever program you're using?

The default tends to be 'you are a great assistant and are helpful to the user' or something. There's other definitions. You can even tell it to be a lazy catgirl maid who meows a lot, and the personality and output completely changes.

the closing was unwanted

i'm just using oobabooga and tried a few things just with instruct and then chat, with chat i use my own template of a chartackter i have made and it works fine, i just want to start getting into making my own loras and something that is a bit tailored towards myself

Mhmm... Yeah that's where i started too. Moved to KoboldCPP which i think is a bit better, along with SillyTavern. Even KoboldCPP's default interface does okay if you don't care much about logs.

Within OobaBooga there's a tab to handle character profiles, which accepts SillyTavern png files. Save the image (it contains a json definition for a bot, you can of course also just get the raw json file too) and import it. There's plenty of places that have character/RP cards, but basic changed assistants can't hurt.

source: https://characterhub.org/characters/ShyChu/aster-a-i-assistant-c91a400eaeb5

You can also with an LLM tell it the tone and feeling and what you want, and then have it help you write a definition too. Though some models tend to be way too terse, don't follow instructions or like huihui models tend to try and be too poetic and flowery in their short 10-20 word 10 line responses. I've found Gemma4 models to be very impressive.