NSQL 350M

The NSQL 350M model, developed by Numbers Station, is an open-source large language model tailored for crafting SQL queries from natural language inputs. It belongs to a model family that also includes larger iterations like NSQL 2B and NSQL 6B and is based on Salesforce's CodeGen models. This autoregressive model constructs queries token by token and is recognized for its ability to generate SELECT queries effectively when provided with a table schema and clear instructions. NSQL 350M undergoes initial training using a corpus of general SQL queries and is fine-tuned on a vast dataset consisting of text-to-SQL pairs drawn from diverse public sources. However, while it excels at generating accurate SQL code for moderately complex queries, its efficacy relies heavily on the prompt structure and may falter with highly intricate queries or those outside its training scope. Despite showing strong potential among open-source models, it is yet to match the precision of proprietary models like GPT-4 in tackling complex tasks.

NSQL 350M is a model in the NSQL family. This page tracks provider routes through Baseten API. No headline benchmark score is tracked for NSQL 350M yet.