danielpedrozo/sql-zero-3b-gold-control
TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold
The danielpedrozo/sql-zero-3b-gold-control is a 3.1 billion parameter model developed by danielpedrozo, part of the SQL-Zero project for Text-to-SQL tasks. It is specifically designed for zero-annotation self-play, adapting the Dr. Zero protocol. This model excels in converting natural language to SQL queries, achieving notable execution accuracy on benchmarks like Spider, BIRD, and Syn.
Loading preview...
Model Overview
The danielpedrozo/sql-zero-3b-gold-control is a 3.1 billion parameter model focused on Text-to-SQL conversion. It is a component of the larger SQL-Zero project, which implements a self-play, zero-annotation approach adapted from the Dr. Zero methodology.
Key Capabilities
- Text-to-SQL Conversion: Translates natural language queries into executable SQL commands.
- Zero-Annotation Training: Utilizes a self-play mechanism, reducing the need for extensive human-annotated datasets.
- Evaluation Protocol: Assessed using greedy@1 with full schema (DDL + 3 rows) and execution accuracy.
- Performance Metrics: Achieves 0.731 on Spider, 0.357 on BIRD, and 0.622 on Syn benchmarks, specifically on 8,659 human pairs from the Spider train set with consistent GRPO.
- Reward System: Trained with
verl(GRPO) and a 3-tier execution reward (1.0/0.1/0.0).
Good For
- Applications requiring robust and accurate natural language to SQL capabilities.
- Scenarios where data annotation is limited or costly.
- Developers interested in models trained with advanced self-play and reward mechanisms for structured query generation.