AgentSkillsCN

manage-datasets

将数据集作为 WandB 工件上传、下载、列出并进行版本管理，以确保可复现性。

SKILL.md
--- frontmatter
name: manage-datasets
description: Upload, download, list, and version datasets as WandB artifacts for reproducibility
disable-model-invocation: true
allowed-tools: Bash(python *)
argument-hint: [upload|download|list|info]

Manage Datasets

Manage WandB dataset artifacts for full reproducibility.

Commands

Upload a dataset

bash
python .claude/skills/manage-datasets/manage_datasets.py --project <project> upload \
  --name <artifact-name> --path <local-path> \
  [--description "..."] [--metadata '{"num_samples": 10000}'] [--alias "train-v1"]

Download a dataset

bash
python .claude/skills/manage-datasets/manage_datasets.py --project <project> download \
  --artifact <name:version> [--output ./data/]

List all datasets

bash
python .claude/skills/manage-datasets/manage_datasets.py --project <project> list

Show artifact info

bash
python .claude/skills/manage-datasets/manage_datasets.py --project <project> info \
  --artifact <name:version>

When to use

  • •Before starting experiments: upload training data for reproducibility
  • •In misc/ scripts: fetch external data, process it, upload as artifact
  • •In main.py / src/: download artifact to use during training
  • •To audit which dataset version was used for a run

$ARGUMENTS