Documentation

1. Upload a dataset. Start with a clean CSV (one row per example). Laplace profiles it, flags issues, and previews the columns.
2. Describe your goal in plain language, e.g. “predict churn from these columns” or “forecast weekly demand.” Laplace routes it to the right task family.
3. Review the plan. Laplace shows the chosen approach, validation strategy, the cost estimate (your Compute Passport), and any warnings (leakage, small samples, imbalance) before training.
4. Train & evaluate. Candidates are trained and compared against baselines; you get metrics and a trust verdict.
5. Take the artifacts: a report (PDF/Markdown), reproducible notebook, model card, prediction script, citation, and a deployment endpoint — all included in the beta.

No sample data yet? Use the in-app sample datasets on the new-project screen to try a full run end-to-end.

Large datasets upload best as a single .zip — zipping a folder lets Laplace stream the archive straight to object storage. If you drop a loose folder that is too large, Laplace will ask you to zip it and try again.

Laplace is a closed beta. New accounts join with an invite code tied to a lab — your code attaches you to your lab’s workspace and shared compute. If you received a code by email, the link drops you on sign-up with it pre-filled; otherwise paste it on the sign-up screen.

On your first run you’ll accept a short Beta Test Agreement (the product is pre-1.0, evolving fast, and your feedback shapes it). Need codes for collaborators, or ran out? Ask your lab’s champion or reach the team.

Instead of plans and credits, every beta lab gets a Compute Passport: a monthly budget measured in internal compute dollars (hardware-neutral, so you never reason about GPU-hours). CPU runs are nearly free; GPU jobs are metered against the budget. Before any GPU run, Laplace shows a cost-and-runtime estimate, and you can watch your balance on the project’s Compute Passport panel and on your account page.

Need more than the self-serve budget allows — a long fine-tune, a big sweep, or a top-tier GPU? Request a moonshot from the run screen; we review it and fund the voucher. When the LLM-assisted planner is unavailable, Laplace falls back to a rule-based parser at no extra cost.

GPU jobs are metered per run and capped by a per-run cost limit and your monthly Compute Passport budget. If a run would exceed either cap, Laplace blocks it with a clear message before any spend occurs, and warns you as you approach the limit. An administrator kill-switch can pause all GPU jobs during incidents. Heavy vision and deep-learning tasks (object detection, segmentation, transformer fine-tunes) run on the GPU worker — see the capability matrix for which tasks require it.

Every model ships with a plain-English trust verdict backed by concrete checks: leakage detection, baseline comparison, group/time-aware validation, small-sample and class-imbalance warnings, out-of-distribution checks, and conformal uncertainty where applicable.

• High, passes the checks with a clear margin over baselines on appropriate validation.
• Medium, usable, but with caveats (limited data, weaker margin, or assumptions to verify).
• Low, treat as exploratory; the data or signal doesn’t yet support a dependable model.

A trust score describes the model on your data; it is not a guarantee. For advisory tasks like causal inference, the verdict cannot validate the underlying assumptions, those remain your responsibility. See the disclaimers on the capability matrix.

Laplace exposes a REST API for projects, datasets, runs, models, and deployments. Authenticate with a bearer token from your account, then call the JSON endpoints (for example, list capabilities at GET /capabilities, or your compute usage at GET /compute/wallet). The API and deployment endpoints are open to every beta lab. Full per-endpoint reference is rolling out — tell us what you’re building via contact.

You own your data. Export your datasets, models, and reports at any time, and delete individual projects, your account, or (for owners) your whole organization from settings. After deletion, content is removed from primary systems within 30 days and from backups within 90 days. See the Privacy Policy and Security Overview.