Gemma4-12b-it-asimov

This is the complete SFT across all modules, using the same dataset that was also used to create the SSFT for comparison. I had to reduce the R and Alpha settings slightly for the full SFT.

training:
  max_seq_length: 4096
  num_epochs: 3
  learning_rate: 0.0002
  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
lora:
  lora_r: 32
  lora_alpha: 64
  lora_dropout: 0.05
  target_modules:
    - q_proj
    - k_proj
    - v_proj
    - o_proj
    - gate_proj
    - up_proj
    - down_proj
  use_rslora: false
  use_loftq: false
  finetune_vision_layers: false

You can see a comparison of the output of both models with a test prompt compared to the base here: https://huggingface.co/ewald1976/gemma4-12b-it-styletuned/blob/main/examples.md