defog/sqlcoder-34b-alpha

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Nov 14, 2023License:cc-by-4.0Architecture:Transformer0.2K Open Weights Cold

Defog's SQLCoder-34B-Alpha is a 34 billion parameter language model fine-tuned on a CodeLlama base, specifically designed for converting natural language questions into SQL queries. It significantly outperforms GPT-4 and other popular open-source models on novel SQL generation datasets, demonstrating high accuracy across various query categories. The model is optimized for SQL generation tasks, making it a powerful tool for database interaction and data analysis.

Loading preview...

Defog SQLCoder-34B-Alpha: Natural Language to SQL Generation

Defog's SQLCoder-34B-Alpha is a 34 billion parameter model built upon a CodeLlama base, engineered to excel at transforming natural language questions into precise SQL queries. This model has been rigorously fine-tuned on over 20,000 human-curated questions across 10 distinct schemas, ensuring robust performance on diverse and novel datasets not seen during training.

Key Capabilities & Performance

  • Superior SQL Generation: SQLCoder-34B-Alpha achieves an 84.0% correctness rate on the sql-eval framework, surpassing gpt-4 and gpt-4-turbo (both at 82.5%) on novel datasets.
  • Category-Specific Accuracy: Demonstrates strong performance across various SQL query categories, including group_by (94.3%), order_by (88.6%), and join (82.9%).
  • Open-Source & Commercial Use: The model weights are licensed under CC BY-SA 4.0, permitting commercial use and modification, provided modified weights are open-sourced under the same terms.

Use Cases & Hardware

  • Efficient Database Interaction: Ideal for applications requiring accurate and efficient conversion of user queries into SQL, streamlining data access and analysis.
  • Accessible Deployment: Can be run on consumer GPUs with 20GB+ memory (e.g., RTX 4090, RTX 3090, Apple M2 Pro/Max/Ultra) using 8-bit or 4-bit quantization, or on 4xA10 GPUs with float16 weights.