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

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-sqale is a 2.3 billion parameter Qwen3.5-based agentic text-to-SQL model. It is specifically trained with GRPO (Generative Reinforcement Learning with Policy Optimization) and a length curriculum to explore databases using tools and generate accurate SQL queries. This model excels at complex text-to-SQL tasks, particularly on challenging datasets like SQaLe3 and BIRD, demonstrating strong performance even when schema information is withheld.

Loading preview...

Model Overview

cwolff/qwen3.5-2b-grpo-sqale is a 2.3 billion parameter Qwen3.5-based model designed as an agentic text-to-SQL system. It is trained using Generative Reinforcement Learning with Policy Optimization (GRPO) against real SQLite execution, incorporating a unique length curriculum that reweights the sampling pool towards longer, more complex queries over time. This approach, combined with training on harder datasets including simple, moderate, and hard questions from the SQaLe3 pair, enhances its ability to handle intricate database interactions.

Key Capabilities & Differentiators

  • Agentic Text-to-SQL: Operates as an agent, exploring databases with a suite of tools (list_tables, describe_table, run_query, submit_sql, etc.) before formulating and submitting SQL queries.
  • Enhanced Performance on Challenging Data: Demonstrates significant gains over its untrained base model, particularly on challenging questions from the BIRD benchmark. It performs comparably to an untrained 9B parameter model when schema information is withheld, a regime it was specifically optimized for.
  • Length Curriculum Training: Utilizes a length curriculum to progressively improve performance on longer and more complex SQL queries.
  • Increased Episode Budget: Trained with a larger episode budget of 24,576 tokens, allowing for more extensive exploration and reasoning.

Usage & Performance

This model is intended for use within an agent loop, where it emits single JSON objects representing tool calls. It was evaluated using an agentic harness with 6 tool turns and specific sampling parameters (T=0.6 / top-p 0.95 / top-k 20), scoring by executing against real databases. It achieves an official score of 0.583 on SQaLe3 test (withheld, full schema) and 0.473 on BIRD dev (schema withheld). The model requires a larger context length of 24,576 tokens for optimal performance during serving.

Finetunes
1 models