jastorj/couchmind-v5.7.6.1_qwen_stage_3-cw-19K-16bit

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

jastorj/couchmind-v5.7.6.1_qwen_stage_3-cw-19K-16bit is a 7.6 billion parameter Qwen2.5-Coder-7B-Instruct model fine-tuned by jastorj for Text-to-SQL generation. It specializes in converting natural language questions into syntactically valid Couchbase SQL++ queries, leveraging the NL2SQL++ v5.7.6.1_qwen_stage_3 dataset with code-with-thought reasoning. This model is optimized for accurate SQL++ query generation based on provided database schemas and natural language input, supporting a 32K context length.

Loading preview...

Model Overview

This model, jastorj/couchmind-v5.7.6.1_qwen_stage_3-cw-19K-16bit, is a specialized fine-tune of the Qwen2.5-Coder-7B-Instruct base model, developed by jastorj. It is specifically engineered for Text-to-SQL generation, focusing on the Couchbase SQL++ query language. The model has 7.6 billion parameters and supports a context length of 32,768 tokens.

Key Capabilities

  • Natural Language to SQL++ Conversion: Translates complex natural language questions into precise and syntactically correct Couchbase SQL++ queries.
  • Schema-Aware Query Generation: Utilizes provided database schemas (including bucket, scope, and collection names with properties) to generate highly accurate and contextually relevant queries.
  • Code-with-Thought Reasoning: Fine-tuned on the NL2SQL++ v5.7.6.1_qwen_stage_3 dataset, which incorporates code-with-thought reasoning to enhance its understanding and generation capabilities.
  • Optimized for Couchbase: Designed to work seamlessly with Couchbase data structures, ensuring correct usage of bucket, scope, and collection identifiers.

Training Details

The model was fine-tuned using LoRA (Low-Rank Adaptation) with Unsloth, on a dataset of 3900 examples. It utilizes 16-bit merged weights and was trained for 3 epochs with a learning rate of 1e-06. The training configuration emphasizes efficient resource usage with settings like gradient_accumulation_steps and bf16 precision.

Good for

  • Developers and data analysts working with Couchbase databases who need to quickly generate SQL++ queries from natural language.
  • Building applications that require dynamic SQL++ query construction based on user input.
  • Automating data retrieval and analysis tasks within a Couchbase environment.

Finetunes
1 models