Sign in to view source links and access this dataset
Description
Neurondb's postgresql-llm dataset contains 211,539 rows of PostgreSQL-specific SQL and PL/pgSQL queries for training and evaluating large language models. Each row is a triplet of question, schema, and SQL, with metadata for filtering by difficulty, category, and source. The dataset was last updated on March 7, 2026.
Use Cases
Fine-tuning language models for SQL generation based on the question-schema-SQL triplet structure.
Evaluating model performance on PostgreSQL-specific queries, which constitute 5.7% of the dataset.
Analyzing query difficulty and category distributions using the provided metadata.
Training models to generate PL/pgSQL procedural code, a subset of the dataset's content.
Strengths
Contains 211,539 total rows of structured query data.
Includes 11,998 rows (5.7%) specifically for PostgreSQL, a distinct database system.
Provides metadata for filtering and analysis, such as difficulty and category.
Has a schema fill rate of 82.2%, indicating most rows include schema context.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
The explanation fill rate is 17.8%, suggesting most queries lack accompanying natural language explanations.
Last updated 2026-03-07 06:15:52; freshness should be verified.