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

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The cwolff/qwen3.5-2b-grpo-curriculum is a 2.3 billion parameter Qwen3.5-2B model fine-tuned as a text-to-SQL agent. It is specifically designed to explore databases using tools and generate SQL queries, trained with GRPO against real SQLite execution. This model incorporates a length curriculum and harder data (simple, moderate, hard) from SQaLe3, making it particularly effective for complex text-to-SQL tasks. It demonstrates strong performance in schema-withheld scenarios, trading evenly with larger models 4.5 times its size.

Loading preview...

Qwen3.5-2B GRPO with Length Curriculum

This model is a 2.3 billion parameter Qwen3.5-2B variant, specifically developed as a text-to-SQL agent. It excels at exploring databases with provided tools and generating accurate SQL queries, having been trained using GRPO (Generalized Reinforcement Learning for Program Optimization) against real SQLite execution environments.

Key Differentiators & Capabilities

  • Warm-started Training: Continues from a checkpoint of cwolff/qwen3.5-2b-grpo-sqale, indicating advanced pre-training.
  • Length Curriculum: Employs a unique training approach where the sampling pool is reweighted by gold-query length quantile, progressively shifting focus from shorter to longer queries over 400 steps.
  • Harder Data: Trained on simple, moderate, and hard difficulty levels from a newer SQaLe3 pair, enhancing its robustness.
  • Increased Episode Budget: Utilizes double the episode budget (24,576 tokens) compared to its siblings, allowing for more extensive exploration.
  • Agentic Approach: Designed to work within an agent loop, where it uses tools like list_tables, describe_table, run_query, and submit_sql to interact with the database.
  • Performance: Achieves competitive results on BIRD and SQaLe3 benchmarks, notably performing on par with an untrained 9B parameter model when the schema is withheld (0.326 vs 0.333 on BIRD challenging, schema withheld).

Good For

  • Text-to-SQL Applications: Ideal for scenarios requiring a model to translate natural language questions into SQL queries, especially when database schema exploration is needed.
  • Complex Query Generation: Its training on harder data and with a length curriculum makes it suitable for generating more intricate and longer SQL queries.
  • Resource-Constrained Environments: As a 2.3B parameter model, it offers strong performance in specific text-to-SQL tasks, potentially serving as an efficient alternative to much larger models.

Limitations

  • SQLite Only: Trained and evaluated exclusively on SQLite databases.
  • Agent Loop Requirement: Underperforms significantly without an agent loop for interaction.
  • Spider 2.0 Performance: Shows near-zero performance on Spider 2.0, indicating it's not optimized for that specific benchmark.
  • Over-exploration: Still exhibits some over-exploration, with a mean of 15.1 tool calls for BIRD (schema shown) and 29.1 (schema withheld), and a notable percentage of episodes hitting the token cap.