asmsense: a 228k stripped x86-64 function dataset for naming functions in binaries

luxurycashew · reddit · 2026-10-18

A developer released asmsense, a dataset plus a small model (Qwen3-8B + LoRA) for naming and describing functions in stripped binaries.

Data: 228,177 functions from 341 permissively licensed C projects, compiled by the author at -O0 through -Os so every function has its real name and source; English and Turkish descriptions included (Chinese planned). All builds are clang/Mach-O on macOS so far.
Known gaps: ARM64/RISC-V/MIPS architectures, GCC/ELF builds, human-checked labels, agent trajectories, and evaluation on public benchmarks like BinMetric, REBench, BinSum, DeBinVul.
Caveat: descriptions were generated by teacher LLMs with access to C source and only sample-checked, so small errors are possible.
Next idea: turn the pipeline into an RL environment (compile, strip, name, score via token-level F1), while noting the reward is easy to game with generic names; the author asks for reward-shaping advice and whether RL beats plain fine-tuning.

Code and data are open-sourced on GitHub.

More from coding & agent