arkcli-datasets

ArkCLI Datasets

First read ../arkcli-shared/SKILL.md, and follow the authentication, output, security, Client Preview, and secondary confirmation rules.

Capability Scope

• Create a Dataset and its first V1, or create a new version for an existing Dataset.
• From a local .jsonl file or object mapping data under the same TOS bucket.
• Resolve the exact DatasetSchema based on the base model's fine-tuning configuration and perform server-side validation.
• Query, update, and delete Datasets, query versions, and download TOS objects in a version.
• Produce a stable train finetune create reference for <dataset-id>:<version-id> .

When handling only the fine-tuning task itself, use ../arkcli-train-finetune/SKILL.mdinstead. Do not use the Raw API to simulate existing dataset product commands.

Execution Guards

1. First run arkcli auth status; if authentication fails, recover according to the shared skill.
2. Before execution, read the target leaf command's --help, and rely on the currently installed version.
3. --dataset-format only accepts LLMSft、LLMPretrain、LLMDPO、LLMRL、EmbeddingSft、ImageGenSft、VideoGenSft; do not send display text or other backend SchemaTypes.
4. --local and --tos-uri are mutually exclusive. Local input only accepts .jsonl, at most 20 files; multiple TOS URIs must belong to the same bucket.--tos-uri must be listable by the current identity and contain at least one .jsonl; do not reuse the ds-ds/* staging path from the CLI upload result; instead use the ark/dataset/<dataset-id>/<dataset-version-id>/ path of a registered version or another readable data path.
5. Creating and adding a version will first validate the inputs one by one;ValidateDataset.StorageRawLocation.Storage does not send AccountID, but subsequently the CreateDatasetVersion.Storage of both workflows must send the current account's numeric AccountID。CreateDatasetVersion does not send the optional Name, and the CLI does not provide this parameter either; the official version number is entirely based on the Version in the server response. If any step fails, retain the uploaded objects or the created Dataset, and do not fabricate a rollback.
6. dataset version get <dataset-id> <Vn|dsv-*> --preview reads at most 100 KiB (102400 bytes) of content per object, consistent with the console Dataset preview limit; the version SID can still omit the Dataset ID. Download supports both dataset download ... --version ...and dataset version download <dataset-id> <Vn|dsv-*>, and neither overwrites existing files.
7. Before deletion, display the number of versions, the number of storage paths, and, when available, the number of objects and total bytes, then confirm separately.
8. Create, version create, validate, download, update, and delete support leaf-level --dry-run. Read-only list/get do not use --dry-run；version get --preview is a real read-only query, not a Client Preview.
9. Client Preview only displays an offline plan; it does not upload, does not validate server-side data, and does not replace deletion confirmation.

Volc local upload uses the TOS endpoint of the region corresponding to the current profile, and enables bucket rename.

Workflow

• Create, maintain, validate, download: read references/commands.md。
• Dataset integration with fine-tuning: after completing the Dataset, read ../arkcli-train-finetune/references/create.md。
• Maintainer evaluation: read references/evals.md。

Only load the reference needed for the current task.