What is it?
This model is my attempt at further improving the results of LFM2.5 1.2b instruct finetune, since 1.2b is very little, I decided to work on a heavier, but still somewhat tiny architecture. So why not qwen 3.5 4b? I honestly tested, hell, even made a similar finetune on it, but roleplay wise, gemma 3 felt better to me.
But generally, this is me trying to create a portable model that anybody can run, I am doing my best to make those smaller models do roleplay.
A word on how good is it:
Advantages:
- Better writing
- less slop
- small size
- improved punctuation
Disadvantages:
- Logic is massively improved over LFM 2.5 1.2b tune, but is not ideal
- Due to being a rather small model, it will tend to sometimes repeat itself, altough I did my best to prevent this from happening
A word on the quants for this model:
- BF16- Mostly overkill, however, highest quality
- Q8_0- Amazing quality, near lossless
- Q6_K- High quality, fast
- Q5_K_M- Mid to high quality, small and fast
- Q4_K_M- Mid quality, very small and very fast
Quants are located in the repo, along with safetensors
V2:
v2 is on the way, I am training it on way larger dataset(17M unique tokens of roleplay data). It will improve current results, and I will make sure it will do so without overfitting. I will continue making small roleplay tunes in order to make local roleplay possible for anyone, no matter the hardware.