What data we have, and what is still missing.
Each block is one kind of data. Solid blocks exist and are measured, orange ones are partly done, dashed ones have not been started. The dashed blocks are the roadmap.
Goal
A model that runs on your own machine, names and explains functions in stripped binaries and raw firmware, and shows the evidence for each claim.
Where the data comes from
We compile permissively licensed open-source C projects ourselves and strip the result. The source gives the ground truth: real names, comments, and which bytes belong to which function.
What is public
The x86-64 dataset, its teacher labels, the v6 model and its test outputs are on Hugging Face and GitHub. Firmware data and the agent work are kept private for now.
Evaluation runs which model saw which input on which test set
All runs score name prediction with token F1 on held-out projects. "context" adds the callees' library calls and strings; "+decompile" adds anonymized Ghidra pseudocode; "reasoning" lets the model think before answering. Result files are in sonuc/.
Other public datasets and benchmarks read, not yet used by us
Work we compare against or plan to evaluate on. A large share of recent binary-analysis research is published by groups in China, which is one reason Chinese is on our language list. Sizes are as stated in each paper's abstract or repository.