AI开发平台ModelArts-yaml配置文件参数配置说明:sft_yaml样例模板

时间:2025-01-26 10:50:38

sft_yaml样例模板

### modelmodel_name_or_path: /home/ma-user/ws/tokenizers/Qwen2-72B### methodstage: sftdo_train: truefinetuning_type: fulldeepspeed: examples/deepspeed/ds_z3_config.json### datasetdataset: identity,alpaca_en_demotemplate: qwencutoff_len: 4096packing: truemax_samples: 1000overwrite_cache: truepreprocessing_num_workers: 16### outputoutput_dir: /home/ma-user/ws/tokenizers/Qwen2-72B/sftlogging_steps: 2save_steps: 5000plot_loss: trueoverwrite_output_dir: true### trainper_device_train_batch_size: 1gradient_accumulation_steps: 8learning_rate: 1.0e-5num_train_epochs: 10.0lr_scheduler_type: cosinewarmup_ratio: 0.1fp16: trueddp_timeout: 180000000include_tokens_per_second: trueinclude_num_input_tokens_seen: true
support.huaweicloud.com/bestpractice-modelarts/modelarts_10_3200.html