wideseek

WideSearch-style broad information seeking: each task poses a complex, multi-faceted query whose gold answer is a Markdown table; the taskset is tool-free (the harness supplies web search) and the agent emits the table as its final message. Table tasks are scored by WideSearch item-level (cell) F1 — rows aligned by the dataset's unique_columns primary keys, cells matched normalized with an optional LLM judge for semantic equivalence — while the short-answer depth split is scored by answer equivalence.

Taskset

• Source: RLinf/WideSeek-R1-train-data (width split; also depth / hybrid)
• Size: 20000 tasks

Changelog

• 2026-08-31: Yield task records on demand so bounded evaluations construct only the requested prefix.
• 2026-08-30: Standardized short-answer equivalence on the native ReferenceJudge contract while retaining the hardened message boundary and specialized batched table-cell judge.
• 2026-07-10: Ported to the task-centric verifiers API: rewards and lifecycle hooks live on the Task (a TaskData row + behavior split), and task-facing config knobs (judges, tool/user placement, scoring parameters) moved from --env.taskset.* to --env.taskset.task.*. Requires verifiers>=0.2.0 and Python >=3.11.
• 2026-06-24: Initial release — WideSearch-style table compilation on RLinf's WideSeek-R1 data; item-level F1 reward aligned by unique_columns with an optional LLM judge.