```html

Efficient SQL Generation with Feather-SQL: A New Approach for Compact Language Models

Translating natural language into SQL (NL2SQL) has made considerable progress thanks to large language models (LLMs). However, these models often depend on proprietary systems and significant computational resources, which presents challenges for data privacy and implementation. Compact language models (SLMs), on the other hand, struggle with NL2SQL tasks and often exhibit lower performance and compatibility issues with existing frameworks.

To address these challenges, Feather-SQL was developed, a new, lightweight framework specifically tailored for SLMs. Feather-SQL improves the executability and accuracy of SQL queries through two primary mechanisms: First, through schema pruning and linking, which removes irrelevant information from the database schema and links the relevant parts with the natural language input. Second, through the generation of multiple paths and candidates, which considers various interpretations of the user input and thus increases the probability of a correct SQL query.

Another innovative aspect of Feather-SQL is the so-called "1+1 Model Collaboration Paradigm." This paradigm combines a powerful, generally trained chat model with a specialized, SQL-fine-tuned model. This collaboration leverages the strengths of both models: the chat model contributes its abilities in analytical thinking and language understanding, while the specialized model is responsible for the precise generation of SQL code.

Initial test results with the BIRD benchmark dataset show that Feather-SQL significantly improves the NL2SQL performance of SLMs. Models without fine-tuning achieved a performance increase of about 10%. Particularly noteworthy is that the proposed "1+1 Model Collaboration Paradigm" raises the accuracy limit of SLMs to 54.76%, highlighting the effectiveness of this approach.

Feather-SQL addresses the current challenges in NL2SQL by improving the performance of SLMs while reducing the need for extensive computational resources. By combining schema optimization, multiple-path generation, and the innovative collaboration paradigm, Feather-SQL offers a promising solution for the efficient and privacy-friendly generation of SQL queries from natural language. This development is particularly relevant for companies like Mindverse, which develop AI-powered content tools and customized solutions like chatbots, voicebots, and AI search engines. By integrating frameworks like Feather-SQL, these solutions can be deployed more efficiently and cost-effectively without sacrificing accuracy and performance.

Bibliography: - https://huggingface.co/papers - https://openreview.net/forum?id=xGkxWP2wE4&referrer=%5Bthe%20profile%20of%20Bingsheng%20He%5D(%2Fprofile%3Fid%3D~Bingsheng_He1) - https://aclanthology.org/2025.coling-main.0.pdf - https://www.vldb.org/pvldb/vol17/p2750-fan.pdf - https://coling2025.org/program/main_conference_papers/ - https://arxiv.org/abs/2406.01265 - https://github.com/HKUSTDial/NL2SQL_Handbook - https://arxiv.org/html/2408.05109v1 - https://www.researchgate.net/publication/383574930_The_Dawn_of_Natural_Language_to_SQL_Are_We_Fully_Ready - https://www.mdpi.com/1999-5903/17/1/12 - arxiv:2503.17811 ```