defog/sqlcoder-70b-alpha

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Jan 30, 2024License:cc-by-sa-4.0Architecture:Transformer0.2K Open Weights Cold

Defog's SQLCoder-70B-Alpha is a 69 billion parameter language model, fine-tuned from CodeLlama-70B, specifically designed for natural language to SQL generation. It features a 32768 token context length and demonstrates superior performance compared to generalist models, including GPT-4, on text-to-SQL tasks. This model is optimized for enabling non-technical users to query SQL databases for analytical purposes.

Loading preview...

Overview

Defog's SQLCoder-70B-Alpha is a 69 billion parameter model, fine-tuned from CodeLlama-70B, engineered for natural language to SQL generation. It is developed by Defog, Inc. and is intended to empower non-technical users to interact with SQL databases for data analysis.

Key Capabilities

  • Superior Text-to-SQL Performance: Outperforms generalist models, including GPT-4, in converting natural language queries into SQL. This is validated on the PostgreSQL-based SQL-Eval framework.
  • High Accuracy Across SQL Categories: Achieves high accuracy across various SQL query types, including date (96%), group_by (91.4%), order_by (97.1%), ratio (85.7%), join (97.1%), and where (91.4%) clauses, as demonstrated in internal evaluations.
  • Large Context Window: Features a 32768 token context length, allowing for complex and detailed natural language inputs.

Intended Use Cases

  • Analytics Tool: Designed for non-technical users to understand and query data within their SQL databases.
  • Read-Only Access: Recommended for use with read-only database access to prevent malicious requests, as it has not been trained to reject such inputs.

Limitations

  • Not designed as a database administration tool.
  • Lacks safeguards against malicious requests from users with write access to databases.