asmsense
asmsense is an open-source project that releases a large-scale dataset of x86-64 functions and trains models for automatic naming and description of functions in stripped-symbol binaries.
Timeline (last 90 days)
asmsense is an open-source project that has released a dataset containing 228,177 x86-64 functions and trained a small model for automatic naming and description of stripped functions.
Every function in the asmsense dataset retains its real name and source code context from 341 permissively licensed C projects.
The asmsense dataset covers multiple compilation optimization levels from -O0 to -Os.
Each function in the asmsense dataset comes with descriptions in English and Turkish, with Chinese descriptions planned.
The function descriptions in asmsense were generated by a teacher LLM that can see the C source code, and only a sample was checked by another LLM, without comprehensive review by human experts.
The asmsense model uses Qwen3-8B as the base model and is fine-tuned via LoRA.
The current asmsense data comes only from clang/Mach-O compilation results on macOS and does not yet cover GCC/ELF builds.
The asmsense dataset currently supports only the x86-64 architecture and does not yet include other architectures such as ARM64, RISC-V, and MIPS.
asmsense has not yet been evaluated on public benchmarks such as BinMetric, REBench, BinSum, and DeBinVul.
The code and data of the asmsense dataset are open-sourced on GitHub under a permissive license.
2 more timeline events
All Facts (12)
The above information about asmsense mainly comes from the project author's one-sided statements on Reddit, and no independent third-party verification has been seen.
50%UnverifiedEvery function in the asmsense dataset retains its real name and source code context from 341 permissively licensed C projects.
50%UnverifiedThe asmsense dataset covers multiple compilation optimization levels from -O0 to -Os.
50%UnverifiedEach function in the asmsense dataset comes with descriptions in English and Turkish, with Chinese descriptions planned.
50%UnverifiedThe function descriptions in asmsense were generated by a teacher LLM that can see the C source code, and only a sample was checked by another LLM, without comprehensive review by human experts.
50%UnverifiedThe asmsense model uses Qwen3-8B as the base model and is fine-tuned via LoRA.
50%UnverifiedThe current asmsense data comes only from clang/Mach-O compilation results on macOS and does not yet cover GCC/ELF builds.
50%UnverifiedThe asmsense dataset currently supports only the x86-64 architecture and does not yet include other architectures such as ARM64, RISC-V, and MIPS.
50%Unverifiedasmsense has not yet been evaluated on public benchmarks such as BinMetric, REBench, BinSum, and DeBinVul.
50%UnverifiedThe code and data of the asmsense dataset are open-sourced on GitHub under a permissive license.
50%UnverifiedThe asmsense authors are considering turning the data processing pipeline into a reinforcement learning environment, but worry that the model could game the token-level F1 reward function by generating generic names.
50%Unverifiedasmsense is an open-source project that has released a dataset containing 228,177 x86-64 functions and trained a small model for automatic naming and description of stripped functions.
50%