hananour/EliteReplay-BIRD-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EliteReplay-BIRD-4B is a 4.5 billion parameter Qwen3.5-4B-Base model fine-tuned by hananour for text-to-SQL tasks, specifically on the BIRD dataset. It utilizes the EliteReplay method, which combines reinforcement learning with supervised replay from an archive of verified SQL solutions to maintain solution diversity during training. This model excels at generating accurate and diverse SQL queries from natural language questions, making it suitable for database interaction and data querying applications.

Loading preview...

EliteReplay-BIRD-4B: Text-to-SQL Model

EliteReplay-BIRD-4B is a 4.5 billion parameter model based on Qwen3.5-4B-Base, specifically fine-tuned for text-to-SQL generation on the BIRD dataset. Developed by hananour, this model leverages the innovative EliteReplay method to enhance its performance and solution diversity.

Key Capabilities

  • Accurate SQL Generation: Translates natural language questions into precise SQLite SQL queries.
  • Diverse Solution Preservation: Employs EliteReplay, a technique combining reinforcement learning with supervised replay from an archive of verified SQL solutions, to ensure a wide range of valid SQL outputs.
  • BIRD Dataset Optimization: Specifically trained on 5,000 BIRD questions, with 1,500 verified SQL examples used for replay, optimizing its performance for complex database querying scenarios.
  • Efficient Training: Utilized GRPO with supervised archive replay, trained on 4 NVIDIA A100 80 GB GPUs.

Good for

  • Text-to-SQL Applications: Ideal for systems requiring conversion of natural language into executable SQL queries.
  • Database Interaction: Useful for building interfaces that allow users to query databases using plain English.
  • Complex Data Querying: Particularly effective for scenarios involving the BIRD dataset's intricate database schemas and questions.
  • Research in RL and Text-to-SQL: Demonstrates an advanced training methodology for preserving solution diversity in reinforcement learning for code generation tasks. More details are available in the EliteReplay paper.

Finetunes
1 models