Beyond Code Pairs: Dialogue-Based Data Generation for LLM Code Translation

Researchers have developed an automated pipeline using dual-LLM agents to generate high-quality training data for code translation tasks, particularly in low-resource languages like Fortran and CUDA. The approach produces verified translations with unit tests and multi-turn dialogue datasets, enabling a 7B model to outperform larger proprietary systems with over 56% improvement in functional correctness on C++-to-CUDA translation.

This research addresses a fundamental challenge in applying large language models to specialized programming domains where training data remains scarce. Traditional code translation relies on paired source-target datasets, but the authors demonstrate that dialogue-based generation—capturing the iterative reasoning behind successful translations—produces substantially better outcomes. By incorporating compiler feedback and runtime validation, the pipeline transforms an otherwise intractable data scarcity problem into a solvable one through synthetic data generation.

The breakthrough has broader implications for AI infrastructure development. Fortran remains critical for scientific computing and legacy systems, while CUDA dominates GPU programming. Both domains face talent shortages exacerbated by code modernization demands. The ability to fine-tune smaller open-weight models to exceed proprietary system performance suggests that specialized AI tools no longer require massive scale—instead, they require smart data generation strategies.

For the AI development community, this validates an emerging pattern: synthetic data quality often matters more than quantity. The 7B model's superiority over larger systems indicates that targeted, verified training data creates more efficient models. This has commercial implications for tools targeting enterprise code migration, where functional correctness demands exceed general-purpose LLM capabilities.

The research trajectory points toward customizable AI translation tools for niche programming domains. Enterprises managing legacy codebases could deploy similar pipelines internally, reducing reliance on proprietary solutions. Future work likely extends this approach to additional language pairs and increasingly specialized frameworks, potentially democratizing code modernization across industries.

• →Dialogue-based synthetic dataset generation outperforms traditional code-pair datasets for LLM fine-tuning on specialized translation tasks.
• →A 7B open-weight model fine-tuned on generated data exceeds larger proprietary systems' performance on C++-to-CUDA translation.
• →Unit test verification integrated into data generation ensures functional correctness, achieving 56% improvement in test success rates.
• →The approach addresses data scarcity in low-resource programming domains like Fortran and emerging frameworks through automated pipeline design.
• →Compiler and runtime feedback incorporation enables iterative refinement captured in multi-turn dialogues, improving translation quality.