What is it?
This is my further improvement of my previous gemma 3 4b it finetune, trained on way bigger amounts of unique data(17M tokens), which resulted in way better writing, bringing the model closer to how 8b and even 12b models write, at times.
Details:
Advantages:
- Way better writing than the base model, less slop and enhanced creativity
- keeps good track of the story
- Small, runs fast
- Better understanding of complex emotional topics
Disadvantages:
- 4b model has limited logic, this tune forces all of this logic to work in order to provide the best creative writing possible
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
What I will do next and is v3 possible?
Next I will probably turn my attention back to smaller models, like Qwen 3.5 2b and LFM 1.2b. I also may train a new version on an even bigger dataset, but it might result in overfitting, overall, this model is the peak performance for a 4b model, or at least the best I could do. working with it wasnt easy, but I did my best and feel satisfied with the results.
PS:(Honestly, I wasted too much time on that, there will soon be a WN9 in Limbus so I probably will finetune less)