experiment: gemma3_1b_qlora
seed: 42
dataset_version: "1.0.0"
model_id: google/gemma-3-1b-it
# LoRA target modules must be verified against the loaded Gemma 3 architecture.
# Do not blindly copy names from other transformers.
quantization:
  load_in_4bit: true
  bnb_4bit_quant_type: nf4
  bnb_4bit_use_double_quant: true
  bnb_4bit_compute_dtype: bfloat16
lora:
  r: 16
  lora_alpha: 16
  lora_dropout: 0.05
  bias: none
  task_type: CAUSAL_LM
  # Placeholder — verify after model load on GPU:
  target_modules: null
training:
  num_train_epochs: 3
  learning_rate: 2.0e-4
  max_seq_length: 512
  per_device_train_batch_size: 1
  per_device_eval_batch_size: 1
  gradient_accumulation_steps: 8
  logging_steps: 10
  eval_strategy: steps
  eval_steps: 50
  save_strategy: steps
  save_steps: 100
  warmup_ratio: 0.03
  lr_scheduler_type: cosine
paths:
  train: data/processed/train.jsonl
  validation: data/processed/validation.jsonl
  output_dir: results/gemma_qlora
inference:
  temperature: 0.0
  do_sample: false
  max_new_tokens: 128