Data Prep

Use this Skill for WF4 dataset analysis, subset design, and data-pipeline preparation.

Read First

• Workflow, language, and documentation rules under ../../../.agents/references/
• ./references/dataset-stats.md
• PROJECT_STATE.json, CLAUDE.md, AGENTS.md when present
• docs/Refined_Idea.md
• docs/20_facts/Execution_Contract.md when present
• docs/30_evidence/Dataset_Table.md when present

Required Work

1. Before writing, check existing WF4 docs/artifacts. Archive current Markdown docs under docs/90_legacy/<YYYY-MM-DD>/; record archive_existing_data_docs_or_NOT_RUN.
2. Resolve dataset name, path, remote source, and subset strategy from state, docs/Refined_Idea.md, execution contract, existing Dataset Table, and the user request.
3. If local data is missing or only a remote source is known, run the Dataset Acquisition Gate before final stats.
4. Auto-detect dataset format and task type.
5. Produce canonical stats and reproducible subset strategy. For NVS/3DGS, preserve task-valid scene, resolution, point-cloud, smoke, dehaze/clean, and evaluation-reference needs.
6. Write docs/Dataset_Stats.md and concise docs/30_evidence/Dataset_Table.md as Conclusion Evidence.
7. Write expected config/script artifacts and update PROJECT_STATE.json, CLAUDE.md dataset paths, and stable AGENTS.md pointer when appropriate.

Dataset Acquisition Gate

Do not stop at “dataset missing” as the first response. First perform Remote Repository Selection, then ask for only the missing decision: download/mount choice and target directory, target archive/slice, network/disk approval, or existing local path.

If target path and approval are already unambiguous, proceed without another question. Otherwise ask before large transfers or writing data outside the repo. Record dataset_acquisition_decision_request_or_NOT_RUN.

Gate Evidence must include source URL, target path, command, result, observed bytes/checksum when available, extraction path, and stats follow-up. With multiple Grill/supervisor candidates, try the next executable candidate after logging failures; skip rejected, deferred, or requires_approval unless separately approved.

Remote Repository Selection

Inspect source-native listings before download: API, repository tree, manifest, README, HTTP metadata, or file-list command. For Hugging Face, check dataset API and relevant tree/main/... listings.

Build a small candidate matrix with remote path/archive, content role (smoke, dehaze, clean reference, depth, COLMAP, point cloud, metadata), resolution/layout, size/checksum when known, required/optional/excluded status, and selection rationale. Do not silently fall back to full data or unrelated conditions.

Output Rules

• Use ./references/dataset-stats.md.
• Separate verified dataset facts from inferred properties.
• Keep Dataset_Table.md source-artifact oriented; do not hand-edit .evidence/**.
• Dataset path sync into CLAUDE.md is required WF4 output.
• AGENTS.md should point to CLAUDE.md, not duplicate volatile paths.
• Report Gate ledger for acquisition, stats, configs, pipeline files, guidance, and state writes, including dataset_acquisition_or_NOT_RUN.

Durable Docs Render

temporary drafts.