saadxsalman/SS-350M-SQL-Strict

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Apr 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The SS-350M-SQL-Strict model by Saad Salman is a specialized 350 million parameter LLM built on the LiquidAI LFM2.5-350M architecture, fine-tuned for strict Text-to-SQL translation. It generates only raw SQL code, eliminating conversational filler and Markdown, making it ideal for high-speed, low-latency SQL generation in resource-constrained environments. Optimized with 4-bit QLoRA and Unsloth, it focuses 100% of its learning capacity on SQL syntax through Completion-Only Loss masking. This model excels at converting natural language questions into SQL queries for schemas with up to 20 tables.

Loading preview...

SS-350M-SQL-Strict: Specialized Text-to-SQL Model

SS-350M-SQL-Strict is a highly specialized, lightweight Large Language Model developed by Saad Salman, specifically engineered for Text-to-SQL translation. Built upon the LiquidAI LFM2.5-350M architecture, this 350 million parameter model is designed to produce only raw SQL code, strictly avoiding conversational text, explanations, or Markdown formatting.

Key Capabilities & Features

  • Strict SQL Output: Generates pure SQL queries without any additional text, ideal for direct integration into applications.
  • Optimized Performance: Leverages 4-bit QLoRA and Unsloth for high-speed, low-latency inference, suitable for edge deployment and resource-limited systems.
  • Focused Training: Utilizes Completion-Only Loss masking during fine-tuning, ensuring the model's learning is entirely concentrated on SQL syntax and schema mapping.
  • Robust SQL Generation: Trained on the Gretel Synthetic SQL dataset, covering complex joins, subqueries, and diverse industry domains.
  • ChatML Prompting: Requires a specific ChatML format to ensure strict behavior and prevent hallucinated text.

Ideal Use Cases

  • Automated SQL Generation: Perfect for applications requiring direct, programmatic conversion of natural language into SQL.
  • Resource-Constrained Environments: Its lightweight nature and optimizations make it suitable for deployment where computational resources are limited.
  • Database Interaction: Streamlining user interaction with databases by translating natural language questions into executable SQL queries.

Limitations

  • Best suited for schemas with fewer than 20 tables.
  • Defaults to standard SQL dialect.
  • Does not provide explanations for the generated SQL.

Finetunes
2 models