Skip to main content
Code can read data through a path inside the container only after the dataset is mounted to a development environment or training job. When mounting, select a dataset that is available to the workload, specify a subdirectory as needed, and configure the container mount path.

​
Mount paths and read/write boundaries

Mounting a dataset does not automatically modify scripts or convert data formats. The data paths in training scripts, Notebooks, or service code must match the container mount path.
  • Separate inputs and outputs: Datasets are suitable as input data. We recommend writing training outputs, checkpoints, Notebooks, temporary files, and intermediate results to a writable Volume.
  • Mount a subdirectory as needed: If a dataset contains multiple versions or subsets, you can mount only the portion you need. This reduces the number of path levels and helps code locate the target data.
  • Avoid overwriting critical directories: Avoid mounting to image-provided directories such as /opt, /workspace, or /usr. Also avoid conflicts with other Volume, dataset, or model mount paths.

​
Mount to a development environment

  1. Create a JupyterLab or VS Code Server development environment.
  2. In the dataset mount configuration, select the target dataset and check its data version and source directory. Development environment dataset mount configuration
  3. Select a subdirectory as needed and configure the container mount path.
  4. Submit the creation request.
After the development environment starts, go to the mount path, list the directory contents, and read a small number of samples to confirm that the data is accessible. We recommend writing code, Notebooks, dependency caches, and intermediate results generated during development to a Volume.

​
Mount to a training job

  1. Create a training job.
  2. In the dataset mount configuration, select the target dataset and check its data version and source directory. Training job dataset mount configuration
  3. Select a subdirectory as needed and configure the container mount path.
  4. Submit the creation request.
After the job starts, the training script should read data from the mount path. Before running the full job, check the directory listing, read a few samples, and verify data loading. Store training outputs, checkpoints, and model files in a Volume to retain and reuse them after the job ends.

​
Troubleshoot mount issues

If the dataset cannot be read inside the container, check the following in order:
  1. Check that the target dataset is available to the workload in the current region and cluster.
  2. Check that the region and cluster are consistent with, or accessible to, the development environment or training job.
  3. Check that the subdirectory is specified correctly.
  4. Check that the code’s read path matches the container mount path.
  5. Check that the mount path does not cover an image-provided directory or conflict with another mount path.