asmsense 发布 22.8 万个 stripped x86-64 函数数据集与 Qwen3-8B 命名模型

asmsense is a dataset and a small model (Qwen3-8B + LoRA) for naming and describing functions in stripped binaries. I compiled permissively licensed C projects myself, so every function has its real name and source.

Right now it has 228,177 functions from 341 projects, compiled at -O0 through -Os, with English and Turkish descriptions (Chinese is planned).

What’s missing is listed step by step on one page: https://asmsense.luxurycashew.com/data/
Code and data: https://github.com/krxi/asmsense

In short: other architectures (ARM64, RISC-V, MIPS), GCC/ELF builds (so far everything is clang/Mach-O on macOS), Chinese descriptions, human-checked labels, agent trajectories, and runs on public benchmarks like BinMetric, REBench, BinSum and DeBinVul.

I’m thinking about turning the pipeline into an RL environment: compile, strip, let the model name the function, and score it against the real name with token-level F1. That reward is easy to game, though: a model can collect partial credit with generic names like get_data. If you’ve trained on a fuzzy target like this, how did you shape the reward, and did RL beat plain fine-tuning?

Two other questions:

• Which of the gaps above matters most for a model that has to run offline?
• How would you evaluate descriptions without a human panel?

A note on LLM use: the descriptions were written by teacher LLMs that could see the C source, and only a sample was checked, by another LLM. I also used LLMs while building this and checking my writings to check grammar, so there may be small mistakes; corrections are welcome.

https://preview.redd.it/bplp468vr7wh1.png?width=1822&format=png&auto=webp&s=1b39946680aa8e3df218ccb688551c658b92f741