# SFT Basic Example
# Fine-tune TinyLlama-1.1B on instruction-following data
# Quick to train (2-3 minutes on consumer GPU) — perfect for testing

model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
data:
  path: examples/data/alpaca_tiny.jsonl
  format: alpaca
task: sft
backend: transformers
quantization: null
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
batch_size: 16
gradient_accumulation_steps: 1
num_epochs: 1
learning_rate: 5e-4
lr_scheduler_type: linear
warmup_ratio: 0.05
weight_decay: 0.0
max_seq_length: 512
output_dir: ./output_sft_basic/
seed: 42
logging_steps: 10
save_steps: 50
eval_steps: 50
eval_strategy: steps
load_best_model_at_end: false