ACE-SQL: Supervised Fine-Tuning and Reinforcement Learning Data for Text-to-SQL

ACE-SQL Training Data is a collection of curated data for supervised fine-tuning, reinforcement learning, and empirical-pool tasks. It was created by xiaobing11 and released on the Hugging Face platform. The data is intended for training a shared language-model policy for text-to-SQL tasks.

Use Cases

• Supervised fine-tuning of language models for SQL generation based on the provided SFT data.
• Reinforcement learning for optimizing SQL generation policies based on the RL data.
• Training a schema retriever model to select required database columns based on the described methodology.
• Empirical research on credit assignment and co-optimization for text-to-SQL systems.

Strengths

• Data is specifically curated for the ACE-SQL methodology, which involves co-optimization of schema retrieval and SQL generation.
• Includes multiple data types for training: supervised fine-tuning, reinforcement learning, and an empirical-pool.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Hugging Face (author: xiaobing11)

Collection Method

Curated for the ACE-SQL research project.

Freshness

Last updated 2026-06-15 12:40:36; freshness should be verified.