Defog’s SQLCoder, a 15B parameter model, outperforms major models in converting natural language to SQL queries. It’s trained on over 10,000 human-curated questions and fine-tuned on a base StarCoder model. The model is open-source and can be used for commercial purposes. It’s been tested on an A100 40GB GPU and can be loaded on consumer GPUs with 20GB or more memory. Future plans include training the model on more data and further tuning with Reward Modelling and RLHF.
Read more at GitHub…