cwolff/qwen3.5-2b-grpo-bird

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The cwolff/qwen3.5-2b-grpo-bird is a 2.3 billion parameter Qwen3.5-based model, specifically fine-tuned as a text-to-SQL agent using GRPO-style reinforcement learning. It excels at exploring databases and generating accurate SQL queries, particularly when schema information is initially withheld. This model is optimized for agentic text-to-SQL tasks on SQLite databases, offering strong performance in a compact size.

Loading preview...

What is cwolff/qwen3.5-2b-grpo-bird?

This model is a 2.3 billion parameter Qwen3.5-based language model, uniquely trained as a text-to-SQL agent rather than a single-shot SQL generator. Developed by cwolff, it utilizes GRPO-style Reinforcement Learning (RL) against real SQLite execution, allowing it to explore databases with tools and submit queries based on execution accuracy. It was trained on the BIRD train dataset (simple + moderate).

Key Capabilities & Performance

  • Agentic Text-to-SQL: Designed for multi-turn interactions, where it can use tools like list_tables, describe_table, and run_query to understand the database schema before generating SQL.
  • Reinforcement Learning Advantage: Shows significant performance gains over its base model, particularly when the database schema is initially hidden (a +0.354 gain in official accuracy on BIRD dev with schema withheld).
  • Compact Size: At 2.3B parameters, it offers a smaller, more efficient alternative for text-to-SQL tasks compared to much larger models, though a 27B model still outperforms it in raw accuracy.
  • High Execution Accuracy: Achieves 0.553 official accuracy on BIRD dev with full schema and 0.547 with schema withheld, with generated SQL executing 97-97.7% of the time.

When to Use This Model

  • Resource-constrained environments: Ideal when VRAM, latency, or cost are critical factors, making a 2B model preferable over larger alternatives.
  • Agentic text-to-SQL tasks: Best suited for scenarios requiring an agent to interactively discover database schema and construct queries.
  • SQLite databases: Specifically trained and evaluated on SQLite, making it highly effective for this database type.

Limitations

  • Small model: Will be outperformed by frontier models on complex multi-join queries.
  • SQLite only: Not tested or optimized for other database systems.
  • Agent loop required: Underperforms significantly without the expected agentic prompting protocol.

Finetunes
1 models