DecBench is a benchmarking suite for evaluating decompilers with a three-stage
pipeline: compile -> decompile -> evaluate. C projects are compiled at
several optimization levels, every in-scope function is decompiled by nine
decompiler backends (two of them LLM coding agents that run on a sampled slice —
see below), and the output is scored against the source and the
binary's DWARF ground truth by three metrics: GED (structural correctness —… See the full description on the dataset page:
https://huggingface.co/datasets/noelo-lab/decbench-dataset.