Wikidata Query Logs Dataset

Organic Query Logs

• Raw Organic Query Logs
• Deduplicated Organic Query Logs

Releases

Each release contains LLM-generated natural language questions for the organic query logs along with the resulting dataset and its KGQA exports (WDQL). The latest symlink always points to the most recent release.

• latest (→ 03-09-26)
• 03-09-26
  • Generations of Qwen3-Next 80B A3B, Qwen3.5 27B, Qwen3.5 397B, Qwen3.6 27B and Qwen3.6 35B A3B
  • Dataset built from those generations
  • WDQL
  • WDQL One Per Cluster
• 21-04-26
  • Qwen3-Next 80B A3B and Qwen3.5 27B Generations
  • Dataset built from Qwen3-Next 80B A3B and Qwen3.5 27B Generations
  • WDQL
  • WDQL One Per Cluster
• 02-02-26
  • Qwen3-Next 80B A3B Generations
  • Dataset built from Qwen3-Next 80B A3B Generations
  • WDQL
  • WDQL One Per Cluster

KGQA Case Study

GRISP (Qwen2.5 7B) trained on the existing Wikidata datasets (GRISP_Prev), only on WDQL 02-02-26 (GRISP_WDQL), or on both (GRISP_All), as in the paper. Browse the outputs in the evaluation app. Trained with GRASP commit 1ba2727.

• Benchmark outputs of all three models on SimpleQuestions, LC-QuAD 2.0, WWQ, QALD-7, QALD-10, QAWiki, SPINACH and WDQL, with evaluations and the GPT-4.1 mini judge comparisons of GRISP_Prev and GRISP_WDQL
• Model checkpoints
  • GRISP_Prev
  • GRISP_WDQL
  • GRISP_All

Other Wikidata Datasets

• Wikidata Benchmarks

Existing Wikidata KGQA datasets (SimpleQuestions, LC-QuAD 2.0, WWQ, QALD-7, QALD-10, QAWiki, SPINACH). They are used for the dataset statistics and comparisons in the paper, and in the case study above: their train and validation splits (SimpleQuestions, LC-QuAD 2.0, WWQ, QALD-7, QALD-10) to train GRISP_Prev and GRISP_All, and their test splits for the evaluation.