BTP: A Research Framework for the Mechanical Interpretability of Code Generation Capabilities in Large Language Models

Introduction / Main Floor: BTP: A Research Framework for the Mechanical Interpretability of Code Generation Capabilities in Large Language Models

The BTP project provides a complete toolchain and experimental framework for analyzing and pruning attention heads in large language models, evaluating the interpretability of the model's internal mechanisms on code generation benchmarks such as HumanEval, MBPP, and LiveCodeBench.