A place to start
Reference data helps when the answer depends on a precise fact rather than fluent prose. Use these datasets to check model terminology, structure a payload, evaluate retrieval, and separate an unsupported answer from a formatting error. Compare encodings on your own data before choosing one.
Practical guide
A small JSON Schema contract for a reference table, plus missing-field, extra-field, and semantic validation checks.
Read the guide →Practical guide
Use a judged query set to separate missing evidence, poor ranking, duplicate context, and generation errors in a RAG workflow.
Read the guide →Practical guide
Build a practical evaluation for factual lookup, evidence coverage, abstention, and tool outcomes without trusting one aggregate score.
Read the guide →Practical guide
Real token counts for all 185 YJTOON datasets in JSON, minified JSON, YAML, and TOON: 19% fewer tokens than minified JSON on table-shaped data, slightly more on list-shaped data.
Read the guide →Make cache behavior, retries, and response headers explicit before your endpoint reaches production.
03Choose joins, windows, indexes, and transaction boundaries from the question you need to answer.
04Find files, recover Git work, and write shell commands that handle real filenames.
05Build usable layouts, keyboard interactions, text search, and time handling.
06Connect service signals, background jobs, probability, and experiment design to real decisions.
