Ko-WideSearch: Korean Web Agent Benchmark Questions

228 Korean-language questions designed to benchmark web agents on exhaustive enumeration tasks. Each task asks an agent to fill every attribute cell of a table by exhaustively enumerating a closed set. Gold answers, source URLs, and scoring details are withheld for a leakage-aware evaluation run privately against held-out data.

Use Cases

• Benchmarking Korean web agents on exhaustive enumeration tasks based on the described table-filling objective.
• Testing information extraction systems on structured knowledge completion tasks based on the described closed-set search.
• Evaluating the breadth-search capabilities of AI models based on the benchmark's design for exhaustive attribute filling.

Strengths

• Contains 228 specific benchmark questions.
• Designed for leakage-aware evaluation with privately held gold answers.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Gold answers and source URLs are withheld, preventing direct validation of model outputs.

Provenance

Source

huggingface

Freshness

Last updated 2026-06-22 05:02:04.