Release v1.0.23 · volcengine/ark-cli

Highlights

Added a complete arkcli dataset command family for creating, listing, inspecting, updating, deleting, validating, versioning, previewing, and downloading managed datasets.
Dataset creation and validation support local JSONL files and existing tos:// prefixes, with server-side schema validation and structured dataset/version results.
Fine-tuning jobs can now use stable <dataset-id>:<dataset-version-id> references, repeatable --train-path entries with either multiplier or sample_count, preset datasets, and managed validation datasets.

Added leaf-level --dry-run plans for dataset create, version create, validate, download, update, and delete workflows.
Mapped TOS sources are listed before validation and must contain JSONL objects; model schema and preset compatibility are checked before fine-tuning submission.
Dataset deletion shows affected versions and storage scope before confirmation, object previews are capped at 100 KiB per object, and downloads never overwrite existing files.
Dataset version numbers come from the authoritative server response, keeping concurrent version creation safe.

Published @volcengine/ark-cli@1.0.23 with binaries for the six supported desktop and server targets.
Added and validated the Volc dataset Skill, and updated the fine-tuning Skill for managed datasets and presets.
Revalidated the resolved Volc Skills tree and CN CDN manifest against the immutable v1.0.23 source before GitHub fallback publication.

Breaking wire-contract disclosure: TrainingSet.Datasets[].Multiplier now accepts a fractional JSON number instead of being integer-only. Existing integer payloads remain valid, but strict schema consumers should update their generated type.
No breaking CLI flag, structured-output field, or public-command changes.
Existing fine-tuning dataset flags remain supported; the new combined references and repeatable dataset-path flags are additive.
GitHub remains a fallback; the Volcengine CDN remains the primary binary and Skills source.