dishants0605/qwen2.5-cobol-python-translator

COBOL to Python Translator

This model, dishants0605/qwen2.5-cobol-python-translator, is a specialized 1.5 billion parameter language model fine-tuned from Qwen2.5-Coder-1.5B. Its primary function is to translate legacy COBOL code into idiomatic Python, aiming to preserve the original business logic. Developed using QLoRA via Unsloth, this model addresses the challenge of manually migrating billions of lines of COBOL code in critical sectors like banking, insurance, and government.

Key Capabilities

• Automated COBOL to Python Translation: Provides a first-pass translation of COBOL code into readable Python.
• Business Logic Preservation: Focuses on maintaining the original intent and functionality of the COBOL code in the Python output.
• Efficient Fine-tuning: Utilizes QLoRA (rank 16, alpha 32) on a custom COBOL↔Python dataset, demonstrating significant training loss reduction from 2.14 to 0.13 over 125 steps.

Good for

• Legacy System Modernization: Ideal for organizations looking to automate the initial stages of migrating COBOL-based systems to Python.
• Reducing Manual Translation Effort: Serves as a valuable tool to reduce the cost and error-proneness associated with manual code migration.
• Code Understanding and Refactoring: Can assist developers in understanding COBOL logic by providing a Python equivalent, facilitating further refactoring and development.

Limitations

It's important to note that the model was trained on a relatively small dataset, which may lead to imperfect translations for complex COBOL patterns like REDEFINES and OCCURS. It is intended as a first-pass aid, not a direct production replacement, and currently lacks formal benchmarks.