LLM Model Training

Tab

Parameter

Required

Description

Default value

Fields Setting

model_name

Yes

The name of the model.

N/A

OSS directory for training output

No

If you do not configure this parameter, the default path of the workspace is used.

N/A

data for training

No

The path of the training data. You can specify a path in Object Storage Service (OSS) or MaxCompute. If you use a different component as the input source, you can skip this parameter. For information about the data format, see Example.

N/A

data for evaluation

No

Optional. The path of the evaluation data. You can specify a path in OSS or MaxCompute. If you use a different component as the input source, you can skip this parameter. For information about the data format, see Example.

N/A

Parameters Setting

learning_rate

Yes

The learning rate.

0.00001

total train epochs

Yes

The number of training iterations.

5

batch size per gpu

Yes

The training batch size for each GPU.

4

maximum sequence length

Yes

The maximum length of a sequence.

512

lora_rank

Yes

The rank that is used to manage the size and sparsity in Low-Rank Adaptation (LoRA) method.

64

gradient accumulation steps

Yes

The number of accumulated gradients before making a single weight update.

1

Tuning

GPU

No

Select a GPU-accelerated instance for the training.

N/A

Maximum Running Duration (seconds)

No

The maximum running duration of the training that uses the component.

N/A