I'm still fascinated by the topic of style tuning using the Gryphe method. I have now started calling it SSFT (Supervised Style Fine-Tuning). It's not about making a model "better" at all, but rather about the countless possibilities that arise from this kind of fine-tuning. That's why I also did another run with Gemma4-12b-it and my dataset.

With Gemma, I had to be a bit more aggressive during the training.

The style change was also significant, and Gemma4-12b-it shows that I haven't changed much, if anything, about Gemma's intelligence and approach. However, the tone and expression have completely changed, in my opinion. Perhaps I'm just so caught up in a loop that I'm imagining it.

Training Setup

yaml

training:
  max_seq_length: 4096
  num_epochs: 3
  learning_rate: 0.0004
  batch_size: 2
  gradient_accumulation_steps: 4
  warmup_steps: 10
  max_steps: 0
  save_steps: 200
  eval_steps: 0
  weight_decay: 0.001
  random_seed: 3407
  packing: false
  train_on_completions: true
  gradient_checkpointing: unsloth
  optim: adamw_8bit
  lr_scheduler_type: linear
  vision_image_size: null
lora:
  lora_r: 128
  lora_alpha: 256
  lora_dropout: 0
  target_modules:
    - lm_head
  use_rslora: false
  use_loftq: false
  finetune_vision_layers: false
  finetune_language_layers: false
  finetune_attention_modules: false
  finetune_mlp_modules: false

An example comparison of vanilla and ssft is here: https://huggingface.co/ewald1976/gemma4-12b-it-styletuned/blob/main/examples.md