CSV ➔ Eval Dataset

How It Works

Paste a CSV with a header row: query (or input), response (or output) and expected (or expected_criteria). Each row becomes one JSONL line holding the input, the output and the expectation as you wrote it. It reads what a spreadsheet exports: quoted cells, commas and line breaks inside quotes, empty cells.

It formats; it does not run a model or mark anything pass or fail. Every expectation is labelled contains_phrase whatever it says, so change the type before another tool relies on it.

Using it to check whether a change made things better or worse: Better Than Last Week? Prove It.

Use Cases

• Structuring user execution history CSVs into RAG evaluation tables.
• Generating programmatic validation rubrics from product spreadsheet specs.
• Bootstrapping assertions for multi-agent validation loops.

Planned CLI

Deploy dataset compilation and assertion rubric tests offline inside your local workspace via our planned command: