Every raw generation with its verdict, straight from the Parquet files on Hugging Face. Queries run in your browser with DuckDB-WASM over HTTP range requests: nothing is downloaded beyond the rows and columns your query touches, and there is no server. Pick a rollouts dataset, then a cell, filter, search inside the generated text, click a row for the full completion with its prompt.
This box is prefilled with the SQL for your current dropdown
selection (each Run refreshes it) — tweak it to go beyond the dropdowns. Write any DuckDB SQL;
the short aliases cell (the selected rollouts partition), failures/executions,
themes, and prompts are rewritten to the matching Parquet files (join on
instance_id, sample; quote "primary" in themes). The prefill uses the full
read_parquet(...) paths directly, which run as-is.