Overview

twinkle-ai/twinkle-sqlcoder is a 24 billion parameter model developed by twinkle-ai, specifically fine-tuned for text-to-SQL generation. It is built on the mistralai/Devstral-Small-2505 base model, utilizing the MistralForCausalLM architecture. The model was trained using bf16 precision with a maximum sequence length of 4096 tokens.

Key Capabilities

• SQL Generation: Translates natural language into SQL queries.
• Full-Parameter SFT: Achieves specialized performance through comprehensive supervised fine-tuning.
• Diverse Training: Leverages a merged dataset including spider, bird, synsql-2.5m, wikisql, and synthetic data for robust SQL understanding.

Intended Use Cases

• Text-to-SQL Research: Ideal for academic and experimental work in converting natural language to SQL.
• Benchmarking: Suitable for evaluating and comparing SQL generation performance against other models.

Limitations

• Generated SQL may require validation before production deployment due to potential inaccuracies.
• Performance is influenced by prompt formatting, the quality of schema context provided, and decoding settings.
• Users should assess safety and compliance requirements prior to deployment.