jastorj/couchmind-v5.7.6.1-cw-5K-16bit
jastorj/couchmind-v5.7.6.1-cw-5K-16bit is a 7.6 billion parameter language model fine-tuned from Qwen/Qwen2.5-Coder-7B-Instruct. It specializes in Text-to-SQL generation, specifically for Couchbase SQL++ queries, leveraging the NL2SQL++ v5.7.6.1 dataset with code-with-thought reasoning. This model is optimized to translate natural language questions into syntactically valid SQL++ queries, making it ideal for database interaction and data retrieval tasks.
Loading preview...
Model Overview
This model, jastorj/couchmind-v5.7.6.1-cw-5K-16bit, is a specialized 7.6 billion parameter language model derived from the Qwen/Qwen2.5-Coder-7B-Instruct base. It has been meticulously fine-tuned for the task of Text-to-SQL generation, focusing specifically on Couchbase SQL++ queries.
Key Capabilities
- Natural Language to SQL++ Translation: Converts natural language questions into precise and syntactically valid Couchbase SQL++ queries.
- Schema Awareness: Utilizes provided database schema information (bucket, scope, collection names, field properties) to generate accurate queries.
- Code-with-Thought Reasoning: Fine-tuned on a dataset that incorporates 'code-with-thought' reasoning, enhancing its ability to generate complex and logical SQL queries.
- Optimized for Couchbase: Specifically trained to understand and generate queries for Couchbase's SQL++ dialect.
Training Details
The model was fine-tuned using LoRA (Low-Rank Adaptation) with Unsloth, leveraging the NL2SQL++ v5.7.6.1 dataset. The training involved 25,528 examples, with an additional 1,062 examples used for validation. The weights are merged and quantized to 16-bit, providing a balance of performance and efficiency.
Ideal Use Cases
- Automated SQL++ Query Generation: For applications requiring dynamic SQL++ query creation from user input.
- Database Interaction: Simplifying data retrieval from Couchbase databases for users without SQL++ expertise.
- Developer Tools: Assisting developers in writing complex SQL++ queries more efficiently.