Research map / Large language models / Language, safety and interpretability
WikiDBs: A Large-Scale Corpus Of Relational Databases From Wikidata
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces WikiDBs, an open corpus of 100,000 multi-table relational databases built from Wikidata with GPT-4o-paraphrased names, and runs initial T5/BART/LM+GNN experiments on missing value imputation, table/column name prediction, and column type annotation.
Problem
Representation learning and foundation models for multi-table relational databases are underexplored because openly available relational database corpora are small and synthetic or confidential.
Approach
Build coherent multi-table databases from Wikidata by traversing foreign-key-like item-valued columns and selecting the next table via Sentence-BERT similarity to the database topic, then paraphrase table and column names with GPT-4o to increase diversity.
Key results
- T5-3B reaches 54.19% cell accuracy and 79.41% column accuracy on WikiDBs with original Wikidata names, dropping to 51.37% and 60.60% after GPT-4o paraphrasing.
- On WikiDBs-10k, T5-3B reaches 60.71% cell accuracy and 92.46% column accuracy, while BART+GNN reaches 30.22% and 69.37%.
- Pre-training T5-3B on WikiDBs yields 44% validation accuracy on GitTablesCTA versus 41% for vanilla T5, and faster early convergence (37% vs 25% after 100 steps).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-4o paraphrased table and column names instead of Wikidata names (ablation) | T5-3B on WikiDBs · WikiDBs, column name prediction | token-wise accuracy: 79.41 -> 60.60 | Hurt |
| Scale from T5-base to T5-3B | T5 on WikiDBs · WikiDBs, cell prediction | token-wise accuracy: 46.87 -> 54.19 | Helped |
| Add a GNN to BART | BART on WikiDBs-10k · WikiDBs-10k test set, cell prediction | token-wise accuracy: 24.14 -> 30.22 | Helped |
| Pre-train T5-3B on WikiDBs before GitTablesCTA fine-tuning | T5-3B · GitTablesCTA validation set | accuracy: 41% -> 44% | Helped |
| Add neighboring tables as context (ablation) | T5-3B + GNN table name prediction · WikiDBs, table name prediction | exact match accuracy: 18.00% -> 18.40% | Helped |
| Add neighboring tables as context (ablation) | T5-3B + GNN table name prediction · WikiDBs, table name prediction | token-wise accuracy: 26.07% -> 28.34% | Helped |
Surprising findings
- Paraphrasing names with GPT-4o makes the tasks substantially harder, reducing accuracies that were inflated by generic Wikidata property names.
- Scaling T5 from 770M to 2.8B gives only modest gains relative to the rising computational cost, and the authors report that language models struggle to use relational context.
Limitations
- Wikidata data may be noisy, untruthful, and skewed toward certain communities.
- Tables are small on average (118 rows), and Wikidata lacks data for larger tables.
- LLM-based paraphrasing inherits hallucination and bias risks.
Experimental setup
Datasets: Wikidata JSON dump GitTablesCTA WikiDBs-10k
Benchmarks: WikiDBs WikiDBs-10k GitTablesCTA
Metrics: token-wise accuracy exact match validation accuracy
Baselines: BART BART+GNN T5-base T5-large T5-3B
Builds on: WikiDBs-10k Sentence-BERT T5 BART TaBERT
Closest papers by meaning
- Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024 · similarity 0.86)
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.86)
- ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.85)
- MMTU: A Massive Multi-Task Table Understanding and Reasoning Benchmark (NeurIPS 2025 · similarity 0.85)
- UniTabE: A Universal Pretraining Protocol for Tabular Foundation Model in Data Science (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.