Skip to content
Beta · version 0.8.1 · commands and the JSON format may still change.

Release notes

Tabalyst Generate unifies the main generated file name to insurance-main.csv across all locales (retiring insurance-principal.csv), with symmetric naming in multi-locale mode (insurance-main-en.csv, insurance-main-fr.csv). Enriches the documentation with a dedicated Locales and languages section and a complete column mapping table between English and French headers. The public showcase demo insurance-customers and its interactive report are updated in English.

Tabalyst Generate delivers the complete Insurance V2 benchmark engine, transforming insurance customer generation into an open-source reference testbed for data quality, sampling strategies and machine learning. Every scenario profile automatically produces a ground-truth suite (truth.csv, rules.json, metrics.json, manifest.json). Adds exact-count rarity levels (rare_events), normalized relational database generation with 4 tables (--relational), operational row-order biases (--row-order), missingness mechanisms (missing_mechanisms), temporal drift and schema evolution (schema_drift), supervised fraud detection targets and leakage-free partitions (splits.csv), and technical stress edge cases. Generation is now in English by default (--locales en, insurance-main.csv, English demographics and titles, English showcase demo), with French support (--locales fr). Includes full Python API support and empirical benchmarks in tabalyst-lab.

Tabalyst Sample gains the mixed method: --first-rows, --random-rows and --last-rows keep the start of the file, a random draw of the other rows and the end, without any row kept twice. Tabalyst Generate shows a progress by rows for insurance and crm as it does for computer.

Tabalyst Report counts the combinations of missing values of a dataset (missing_patterns, in the second pass), names the blocks of columns that go missing together in the report, and --details writes a first chart page, charts-missing.html, with an UpSet plot of those combinations. --details now runs the second pass, as --deep does, and the chart pages load Apache ECharts from a CDN (pinned by its SRI hash); report.html still loads nothing from outside. The second-pass document is at revision 5 and the Report profile at revision 6.

Tabalyst Generate ties the columns of its examples together, so the files can test analyses across columns: CRM deal amounts, probabilities and stages, dates in order and sales territories; Insurance travel contracts, claims and premiums by age; Computer reviews that come with time. The clean files hold no exception to these rules, and every example gets a rule_breaks profile that breaks them on purpose, each change named in the change log. The same seed gives different files than with 0.7.4. No JSON format changes.

Tabalyst Sample gets a second stratified method. stratified-balanced keeps up to --per-value N random rows of each value of one or several fields, and all of a smaller value, so that rare values are well represented; its size follows from the data. stratified now keeps at least one row of each value, and stops with a clear error when the size asked for is under the number of values (before, a rare value could get no row, so a sample can differ from one made with 0.7.3). No JSON format changes.

Tabalyst Generate is much faster for Computer (30,000 rows with errors in two languages in about 17 seconds instead of 8 minutes, and faster again with several processes) and easier to use for one dataset. -o writes a single Computer CSV, with errors when you give an anomaly profile (--anomaly-profile); --locales en or --locales fr writes one language; --anomaly-profile takes several profiles to get one variant of each; --workers sets the number of processes. The command shows its progress, with the time left, and ends with a timing report. The same seed still gives the same data, and no JSON format changes.

Report tells the repetitions of an identifier column that exact copies of whole rows explain from those between distinct records (identifier_uniqueness, report --dedup-report), with a non-binding near-unique reading. The second pass can count the copies of rows and the distinct values after normalization (scan --deep). Report moves to format revision 5 and the second pass to revision 4, so older documents must be made again.

Adds Tabalyst Dedup: tabalyst dedup finds the exact duplicate rows of CSV files and writes the unique rows, the repeated rows and a JSON report, with batches, incremental runs and the Python API (dedup_csv, generate_dedups).

Reworks Scan and the JSON formats: samples with stated accuracy, inferred patterns, observations, a Scan web page and a second pass (scan --plan, scan --deep) that Report chooses automatically when it pays. Adds smart samples (sample --sample-method smart). Scan and Report move to format revision 4, so older documents must be made again. Generate now reads package directories instead of definition_version: 1 files and adds the Computer example.

Adds Tabalyst Generate for synthetic CSV datasets from YAML definitions, with packaged CRM and insurance examples and optional logged anomaly variants.

Makes choosing an Excel table or collection easier. Reorganizes the documentation.

Adds support for Excel workbooks.

Chooses a JSON collection from report and from inspect’s own output.

Adds Tabalyst Inspect and JSON Lines support.

Fast default CSV scans.

First beta release.

Makes scans several times faster on large files.

Adds adaptive detection for columns with many distinct values and a report time that includes the scan.

Introduces Tabalyst Scan, a streaming analysis engine for CSV and JSON.

Introduces Tabalyst Sample.

Introduces a command-oriented toolkit: tabalyst report INPUT... [OPTIONS].

First packaged releases.