| 1234567891011121314151617181920 |
- mlflow_project: sft-gsm8k
- checkpoint_dir: ${hydra:run.dir}/checkpoint/
- csv_dir: ${hydra:run.dir}/csv/ # for saving evaluation results
- model_path: huggingface_models/Qwen/Qwen2.5-Math-1.5B
- dataset_path: "data_sft/openai/gsm8k"
- train_device: cuda:0
- eval_device: cuda:1
- micro_batch_size: 2
- gradient_accumulation_steps: 4
- eval_batch_size: 2
- num_epochs: 1
- learning_rate: 2e-5
- weight_decay: 0.01
- eval_steps: 1
- save_steps: 10
|
