Learning Bug Context for PyTorch-to-JAX Translation with LLMs

The article introduces T2J, a benchmark designed to address translation bugs in converting PyTorch code to JAX using large language models (LLMs). It consists of 20 kernels from the TorchLeet dataset, which were translated by the weak LLM gpt-4o-mini and subsequently debugged by software developers. The T2J benchmark was shown to enhance the translation quality, achieving up to a 20% improvement in the T2J-CodeTrans-Score, highlighting its potential utility for practitioners working on code translation tasks between these frameworks.