compute:
  instance_type: A100:32GB
  region: us-east-1  # or your preferred region

base_model: meta-llama/Llama-2-7b-hf
model_id: Lukeam/llama-aa-fine-tuned
dataset: processed_qa.jsonl

training:
  learning_rate: 5e-5
  num_train_epochs: 3
  per_device_train_batch_size: 2
  gradient_accumulation_steps: 4
  save_steps: 1000
  eval_steps: 500