MALWIN

A function-name-recovery benchmark for stripped Linux malware, across four architectures (x86, x64, ARM, MIPS). Each example is a Ghidra-decompiled function body with placeholder names, paired with the author-written names read from the unstripped symbol table.

Dataset of the ACSAC 2026 artifact for R+R: Revisiting LLM-Based Binary Name Recovery for Real-World Malware Analysis. Models trained on it: MALWIN-CodeGemma-2B and MALWIN-Qwen2.5-3B.

Splits

Four subsets (one per architecture), each with four splits. IT = in-training family; NIT = not-in-training, i.e. whole malware families held out from training, which is the split that measures generalisation to unseen families.

Subset	Binaries	train	valid	it_test	nit_test
x86	559	35,026	6,316	1,067	2,601
x64	776	87,925	10,084	1,956	5,023
arm	862	44,172	7,884	1,550	971
mips	692	36,607	7,559	1,632	1,570

from datasets import load_dataset
d = load_dataset("nghi85/MALWIN", "x86")      # or "x64", "arm", "mips"
print(d["nit_test"][0]["query"])

Schema

GenNm's alpaca-style schema, so the data drops into GenNm-compatible training and evaluation code.

column	meaning
prog_name	<family>__<sha256> of the source binary (family = MalwareBazaar signature, lower-case)
strip_func_name	Ghidra placeholder of the prompted function, FUN_<addr>
prompt	empty (body-only prompts)
query	the Ghidra-decompiled pseudo-C body, followed by Q:[...], the placeholders to name (the function itself and the callees it references)
response	ground truth, A:{'FUN_<addr>': '<author name>', ...} for every placeholder in Q:[...]
hints	unused, empty (kept for GenNm compatibility)

Model input is prompt + query; response is parsed with GenNm's parser.

Note: a handful of IT binaries carry the mis-spelled family label gafygt; the paper's scoring code treats it as gafgyt.

How it was built

• Binaries routed by ELF header into the four architectures (32-bit x86 and x86-64 kept separate).
• Ground truth: names from the unstripped symbol table, joined by address with the Ghidra 11.2.1 decompilation of a stripped copy, so the model never sees a name it is asked to predict.
• Library and runtime symbols removed with a curated classifier, leaving author-written functions.
• NIT test = whole families held out. The remaining families are split 70/10/20 by function count into train / valid / IT test.
• Leakage filter: any IT/NIT test function whose body exceeds a fuzzy-match ratio of 90 against any training function is dropped. Train and valid are never filtered.

The full construction pipeline, SHA-256 manifests and verification tools ship in the artifact under artifact/code/malwin_build/.

Provenance, ethics and safe handling

The 2,889 source binaries are public malware samples from MalwareBazaar (abuse.ch) that shipped with an intact symbol table.

No binaries and no executable code are redistributed. This dataset contains Ghidra-decompiled pseudo-C text (from stripped copies, so functions appear as FUN_<addr>) and the author-given function names from the symbol tables. Binary identities are SHA-256 hashes only.

Decompiled malware text may contain strings such as C2 hostnames, credentials, or offensive words written by malware authors. Treat these fields as untrusted text: nothing here is meant to be executed, and the artifact never executes any of it. No human-subject data and no personal data of victims is included.

License

CC BY 4.0. If you use MALWIN, please cite R+R: Revisiting LLM-Based Binary Name Recovery for Real-World Malware Analysis (ACSAC 2026).