Walkthroughs
Click-by-click guides that work end-to-end on the live product — each one ships pre-configured state and is verified against a real run, with real screenshots. No "works in the docs, breaks in the product."
New to the platform? Start with W1 — it takes ~3 minutes.
Every walkthrough states the role/tier it was verified with and what it costs. Screenshots capture the real console for the account type shown.
The guides
W1 · Upload your first dataset
Upload a CSV into your managed S3 and browse + preview it in the Files view. ~3 min · free · Editor role · no prerequisites
W2 · Import a HuggingFace dataset
Pull a public HuggingFace dataset into hf-datasets/ as JSONL — ready for post-train
(DPO/ORPO, SFT, chat) and GRPO.
~3 min · free · Editor role
W8 · Clean and order a raw CSV with a Data Prep pipeline
Build a Data Prep pipeline in the UI — read data/county-wages.csv from managed S3,
remove duplicates (county + year) and sort by year, write the result to a
managed-S3 output folder, and run it on the Serverless engine.
~10 min · free (Serverless DuckDB) · Editor role
W9 · Ask AI a question about your data
Ask a natural-language question and get a live SQL-generated answer with a result table you can export. ~3 min · ~$0.01 in credits · Viewer role is enough
How walkthroughs are verified
Each guide lives as a Scenario Pack beside the product code
(data-catalog-platform-mvp/walkthroughs/): scenario.json is the single source of
truth, the README you're reading is rendered from it, and a pack check
(npm run test:walkthroughs) fails if the doc drifts from the code it documents.
Walkthroughs are exercised against the live product before they're published —
several of them have caught real bugs (for example, W9 surfaced a crash on count
queries that was fixed and shipped).
More guides are on the way: fine-tuning, post-training, NN training from Parquet, deploy + monitor, and more.