Shirdel-Agent-4B
Agent/Programming models were designed to solve mathematical problems by integrating text-based reasoning with code blocks executed by Python interpreter. The models were trained on python-edu-cleaned, a math instruction tuning dataset with 7.6M problem-solutions:
- Best performance on tasks that resemble the training data
- Better Python explanations
- Calibrated idiomatic Python code
- Better handling of educational programming prompts
- Improved debugging advice
- More consistent code formatting
We did not reduce the quality of the model and for some reason did not use quantization for code quality.
The pipeline we used to produce these models is fully open-sourced.