User Guides

Data Preparation

Prepare tabular datasets and media manifests for successful uploads.

Quickstart
Get from sign-in to a completed first job.
Choosing a Service
Match your data and goal to the right workflow.
Data Preparation
Prepare tabular datasets and media manifests for successful uploads.
Jobs and Results
Understand statuses, reports, SHAP, and outputs.
Billing and Credits
How charges, estimates, and refunds work.
Prediction Guide
Use trained jobs for manual and batch prediction.
Troubleshooting
Resolve common user-facing errors quickly.
FAQ
Short answers to the most common platform questions.
Reference
Handbook and Technical Reference

Open the handbook for the full product guide, then use the model-method appendix when you need mathematical details.

Open Handbook Open model-method appendix

Data Preparation

Supported file formats

  • Delimited text: .csv, .tsv, .tab, .txt
  • Spreadsheets: .xlsx, .xls
  • Columnar files: .parquet, .feather
  • JSON tables: .json, .jsonl, .ndjson
  • Pickled pandas objects are not accepted for user uploads because Python pickle deserialization is unsafe for untrusted files.
  • Statistical data files: Stata .dta, SPSS .sav, .zsav, .por, SAS .sas7bdat, .xpt, and R .rds, .rda, .rdata when the deployment has the relevant optional readers installed

Image, audio, and video workflows are separate from statistical dataset uploads. Use media files or ZIP bundles plus a tabular label manifest in any supported structured-data format. Canonical names are filename and target; common aliases such as path, file_path, video_name, label, class, and tag are accepted.

General preparation rules

  • Use readable, unique column names
  • Keep one row per record unless the workflow is longitudinal or time series
  • Remove duplicate columns and empty trailing rows
  • Make missing values explicit instead of leaving malformed mixed cells
  • Avoid merged cells, decorative title rows, and multi-line spreadsheet headers

Prepare for the workflow you actually chose

Classification

  • use a categorical target
  • check class imbalance before training
  • clean inconsistent label spellings such as Yes, yes, and Y

Regression

  • use a numeric target
  • inspect skew and extreme outliers
  • confirm the target is not accidentally stored as text

Count regression

  • use a nonnegative count target
  • check zero inflation and sparse outcomes
  • confirm the target is not a rate or percentage disguised as a count

Longitudinal workflows

  • include a subject identifier
  • include a valid time column
  • confirm repeated observations exist for the same subject

Time-series workflows

  • use a properly ordered time column
  • remove duplicate timestamps unless aggregation is intended
  • inspect gaps, irregular intervals, and time-zone inconsistencies

Anomaly detection and clustering

  • remove identifier-only columns
  • confirm the feature space contains real behavioral or measurement variables
  • scale-sensitive workflows benefit from cleaner numeric inputs and reduced leakage

Survival analysis

  • include event time and event indicator columns
  • confirm survival times are positive
  • confirm the event flag uses a consistent encoding

Recommended preparation checklist

  1. Confirm the file type and sheet structure are simple enough to parse cleanly
  2. Confirm required columns match the selected service
  3. Remove obvious leakage columns, such as post-outcome fields
  4. Check missingness burden and mixed-type columns
  5. Standardize dates, labels, and units
  6. Keep a copy of the exact uploaded file for reproducibility

Before upload

  • download the workflow template when available
  • read the required columns shown on the upload page
  • start with a comparison strategy, then revise the visible model catalog if needed
  • check that the target column selected in the UI matches the dataset you prepared

Common upload problems

  • missing target column
  • wrong file type or corrupted sheet structure
  • mixed numeric and text values in the same column
  • duplicated timestamps or subject-time pairs
  • identifier columns accidentally treated as predictive features
  • label spelling drift across the same target

A good first upload looks like

  • clean headers
  • one consistent row structure
  • no decorative formatting
  • obvious target and feature columns
  • the simplest version of the dataset needed for the first baseline

Related guides

  • Choosing a Service
  • Quickstart
  • Troubleshooting