Platform For AI:Machine reading comprehension training

Limitations

Runs only on DLC compute resources.

Parameters

• Input ports

  Input port (from left to right)	Data Type Restrictions	Recommended upstream component	Required
  Training data input	OSS	Read OSS Data	Yes
  Validation data input	OSS	Read OSS Data	Yes

• Component parameters

  Tab	Parameter	Description
  Fields setting	Select language	Language of input file. Supported languages: zh (default) en
  	Input data format	Data format of each column in input file. Use commas (,) to separate columns. Default value: qas_id:str:1,context_text:str:1,question_text:str:1,answer_text:str:1,start_position_character:str:1,title:str:1.
  	Question column	Column name containing questions in input file. Default value: question_text.
  	Context column	Column name containing context text in input file. Default value: context_text.
  	Answer column	Column name containing answers in input file. Default value: answer_text.
  	ID column	Column name containing IDs in input file. Default value: qas_id.
  	Start position column	Column name containing start position of answers within context text in input file. Default value: start_position_character.
  	Model save path	OSS bucket folder path to store model files generated after training or fine-tuning.
  Parameters setting	Batch size	Batch size for training. Integer. Default value: 4. For multi-GPU servers, specifies batch size per GPU.
  	Maximum context length	Maximum length of context for processing. Integer. Default value: 384.
  	Maximum question length	Maximum length of questions for processing. Integer. Default value: 64.
  	Sliding window size	Size of the sliding window used to chunk context. Integer. Default value: 128.
  	Number of epochs	Total number of training epochs. Integer. Default value: 3.
  	Learning rate	Learning rate for model building. Float. Default value: 3.5e-5.
  	Save checkpoint steps	Number of training steps after which model is evaluated and best-performing model is saved. Integer. Default value: 600.
  	Select model	Path of pre-trained model provided by the system. Valid values: Custom hfl/macbert-base-zh (default) hfl/macbert-large-zh bert-base-uncased bert-large-uncased
  	Custom model path	Available when Select model is set to Custom. To use a custom pre-trained or fine-tuned model, specify its path here. Format: {A: xxx, B: xxx}. Use colons (:) to separate keys and values. Use commas (,) to separate multiple parameters.
  Execution tuning	GPU instance type	GPU instance type of compute resource. Default value: gn5-c8g1.2xlarge (8 CPU cores, 80 GB memory, single P100 card).
  	Number of GPUs per worker	Number of GPUs per worker. Default value: 1.

• Output port

  Output port (from left to right)	Data type	Downstream component
  Model save path	OSS path. This is the OSS path specified for Model save path parameter on Fields setting tab. Trained model is stored in this path.	Machine Reading Comprehension Prediction

Example

Build a workflow using this component:

Configure the component:

1. Prepare training dataset and validation dataset, and upload them to an OSS bucket. For more information, see Step 2: Upload a file.

   Datasets can be TSV or TXT files and must include the following columns:

   • Training dataset

     ID column, context column, question column, answer column, start position column, and title column (optional)

   • Validation dataset

     ID column, context column, question column, answer column (optional), start position column (optional), and title column (optional)

   This example uses TSV files to demonstrate model training.

2. Use Read OSS Data-1 and Read OSS Data-2 components to read training and validation datasets. Set OSS Data Path parameter to the OSS path where datasets are stored.

3. Connect training and validation datasets to Machine Reading Comprehension Training component and configure its parameters. For more information, see Component parameters in this topic.

References

• Connect Machine Reading Comprehension Prediction component downstream of Machine Reading Comprehension Training component to perform offline prediction on generated model. For more information, see Machine Reading Comprehension Prediction.

• For more information about Designer components, see Designer overview.

• Designer provides various algorithm components. Select appropriate components for data processing based on your scenario. For more information, see Designer component overview.