DEx-3B-GGUF
model for data engineering and data processing
2 min readupdated
DEx-3B is a compact 3-billion-parameter language model fine-tuned from Qwen/Qwen2.5-Coder-3B-Instruct specifically for data engineering and Data Science tasks. It helps write and debug complex SQL queries, Python code (including modern Pandas and Polars 2.x libraries), and dbt models. The model comes in ready-to-use GGUF formats for running locally on regular hardware or Apple Silicon. It is well-suited for data analysts and engineers who need a fast local assistant for writing queries and pipelines.
Key features
- Trained on 15,598 verified examples from real-world datasets and SQL scripts.
- Significantly outperforms the base model on SQL and ETL tasks thanks to a specialized dataset.
- Available in multiple GGUF quantizations (from lightweight Q4 to full F16).
- Fully ready for use in Ollama, llama.cpp, and LM Studio.
What you can do
- Write complex SQL queries with multiple joins, window functions, and CTEs.
- Create data processing scripts in Polars 2.x, Pandas, or PySpark.
- Set up and refactor dbt models, as well as write YAML tags.
- Debug pipelines and troubleshoot Spark crashes or Airflow errors.
Example
ollama create dex-3b -f Modelfile
ollama run dex-3bRead more
DetailsBenchmarks: overall Pass@1 increased from 66.7% to 86.7% compared to the base model
SQL Pass@1 improved from 50.0% to 75.0% on unseen schemas
Hardware requirements: from ~1.8 GB VRAM/RAM for the Q4_K_M quantization
Limitations: at 3B parameters, complex multi-file repositories exceed the model's capabilities










Comments