AIJian/TrustSQL-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AIJian/TrustSQL-8B is an 8.2 billion parameter Text-to-SQL model fine-tuned from Qwen3-8B, developed by Ai Jian and collaborators. It specializes in generating SQL queries from natural language over unknown database schemas, utilizing multi-turn reinforcement learning and tool integration. The model achieves 65.8 EX on BIRD-Dev with greedy decoding and is designed for robust SQL generation within a tool-using agent loop to prevent schema hallucination.

Loading preview...

TrustSQL-8B: Advanced Text-to-SQL for Unknown Schemas

TrustSQL-8B is an 8.2 billion parameter language model developed by Ai Jian and collaborators, specifically fine-tuned for Text-to-SQL tasks. Built upon the Qwen3-8B architecture, this model excels in generating accurate SQL queries even when the database schema is initially unknown. It leverages a novel approach combining multi-turn reinforcement learning with tool integration, designed to enhance reliability and reduce common issues like schema hallucination.

Key Capabilities and Features

  • Specialized Text-to-SQL: Optimized for converting natural language questions into SQL queries, particularly in complex scenarios where database schemas are not fully known beforehand.
  • Tool-Integrated Reinforcement Learning: Employs a unique two-stage training pipeline involving SFT warm-up and Phase-Aware GRPO optimization, guided by an Explore → Propose → Generate → Confirm interaction protocol.
  • Schema Hallucination Prevention: The explicit 'proposal checkpoint' within its inference setup is crucial for verifying tables and columns via tool output, significantly mitigating the risk of generating SQL based on incorrect schema assumptions.
  • Performance Benchmarks: Achieves competitive results on challenging Text-to-SQL datasets, including 65.8 EX on BIRD-Dev (greedy decoding) and 83.9 EX on Spider-Test, demonstrating its effectiveness in real-world scenarios.
  • Robust Inference Workflow: Designed to operate within an agent loop that explores metadata, proposes verified elements, generates and executes SQL, and confirms results, allowing for iterative refinement.

Ideal Use Cases

  • Database Query Automation: Automating the generation of SQL queries from natural language requests for databases with potentially evolving or unfamiliar schemas.
  • Business Intelligence Tools: Integrating natural language interfaces into BI platforms, allowing users to query data without deep SQL knowledge.
  • Data Exploration: Assisting data analysts and scientists in exploring new datasets by translating their questions into executable SQL, even when schema details are not fully memorized.

Limitations: The model was primarily trained and evaluated using SQLite-based benchmarks. Its inference requires a live and secure metadata/execution environment, and generated SQL should always be validated, preferably with read-only permissions initially.

Finetunes
1 models