Volumes

​

Overview

• One file as a dataset. datasets.from_volume(volume, path) connects a file inside the volume as a dataset. The file is downloaded to the training machine when the job starts, like an uploaded dataset. This works today for every provider.
• The whole volume mounted. A job-level volume copies the entire volume to /mnt/<name> on the training machine before training starts, read-only. Custom scripts see it at the same path and in VERI_VOLUME_DIR. This is coming soon: until it is enabled, a job with volume set is refused at submit with a message that points at from_volume.

​

When to use a Volume

• You expect to run more than one training job against the same dataset.
• Your dataset is larger than a few GB.
• Your dataset isn’t public and isn’t on Hugging Face.
• A custom script needs more than one input file.

​

Billing

​

Quickstart

veri volumes create my-corpus

veri volumes upload my-corpus ./train.jsonl

veri volumes upload-dir my-corpus ./training-data --to /data

from veri_sdk import Client

client = Client()
dataset = client.datasets.from_volume(volume="my-corpus", path="/train.jsonl")
job = client.training_jobs.create(
    base_model="Qwen/Qwen2.5-0.5B-Instruct",
    dataset_id=dataset.id,
    output_name="my-run",
    method="sft_text",
    gpu_type="L4-24GB",
    gpu_count=1,
)

​

Mounting a whole volume

job = client.training_jobs.create(
    job_type="custom_script",
    script={"base_image": "...", "code_artifact_id": "...", "entrypoint": "python train.py"},
    volume="my-corpus",
    output_name="my-run",
    gpu_type="L4-24GB",
    gpu_count=1,
)

1. Asks the control plane for the volume’s file list, in pages, with a short-lived download link per file. No cloud credentials reach the machine.
2. Checks that the volume fits on the local disk. If it would use more than 80% of the free space, the job fails right away with volume_too_large_for_node instead of filling the disk mid-run. Local disk is about 150 GB on AWS, 200 GB on GCP, 100 GB on Vast and 720 GiB on DigitalOcean.
3. Downloads the files 16 at a time into /mnt/<name>, checks every file’s size against the list (a mismatch fails the job with volume_sync_incomplete), and makes the tree read-only.
4. For a custom script, bind-mounts /mnt/<name> into the container read-only and sets VERI_VOLUME_DIR=/mnt/<name>. A managed job whose dataset was connected from the same volume reads the file from the mount.

​

Deletion

veri volumes delete my-corpus

​

API + CLI references