Hugging Face's logo

Datasets:
oleksiihrush
/
ollvm-decompilation-dataset

Tasks:
Other
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
decompilation
obfuscation
ollvm
ida-pro
hex-rays
reverse-engineering
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
program
stringclasses
50 values
source_family
stringclasses
2 values
variant
stringclasses
16 values
passes
stringclasses
8 values
source_code
stringclasses
100 values
decompiled_code
stringlengths
976
3.39M
source_lines
int64
16
329
decompiled_lines
int64
48
25.8k
attendance_monitoring
original
build_plain
#include <stdio.h> #include <stdlib.h> #include <string.h> // Define the maximum number of records and length of names #define MAX_RECORDS 100 #define MAX_NAME_LEN 50 // Define the structure to store attendance records struct AttendanceRecord { int id; char name[MAX_NAME_LEN]; int isPresent; }; // Declar...
// Decompiled with IDA Pro Hex-Rays // Input: /Volumes/Personal/University/s2026/diploma/dataset/build_plain/attendance_monitoring // ───────────────────────────────────── // Function: _addRecord // Address : 0x100000550 // ───────────────────────────────────── __int64 addRecord() { __int64 v1; // x8 _BYTE v2[60]...
115
207
attendance_monitoring
original
build_sub
sub
#include <stdio.h> #include <stdlib.h> #include <string.h> // Define the maximum number of records and length of names #define MAX_RECORDS 100 #define MAX_NAME_LEN 50 // Define the structure to store attendance records struct AttendanceRecord { int id; char name[MAX_NAME_LEN]; int isPresent; }; // Declar...
// Decompiled with IDA Pro Hex-Rays // Input: /Volumes/Personal/University/s2026/diploma/dataset/build_sub/attendance_monitoring // ───────────────────────────────────── // Function: _addRecord // Address : 0x100000550 // ───────────────────────────────────── __int64 addRecord() { __int64 v1; // x8 _BYTE v2[60]; ...
115
207
attendance_monitoring
original
build_fla
fla
#include <stdio.h> #include <stdlib.h> #include <string.h> // Define the maximum number of records and length of names #define MAX_RECORDS 100 #define MAX_NAME_LEN 50 // Define the structure to store attendance records struct AttendanceRecord { int id; char name[MAX_NAME_LEN]; int isPresent; }; // Declar...
// Decompiled with IDA Pro Hex-Rays // Input: /Volumes/Personal/University/s2026/diploma/dataset/build_fla/attendance_monitoring // ───────────────────────────────────── // Function: _addRecord // Address : 0x100000550 // ───────────────────────────────────── __int64 addRecord() { int v0; // w8 int v1; // w8 in...
115
1,433
attendance_monitoring
original
build_bcf
bcf
#include <stdio.h> #include <stdlib.h> #include <string.h> // Define the maximum number of records and length of names #define MAX_RECORDS 100 #define MAX_NAME_LEN 50 // Define the structure to store attendance records struct AttendanceRecord { int id; char name[MAX_NAME_LEN]; int isPresent; }; // Declar...
// Decompiled with IDA Pro Hex-Rays // Input: /Volumes/Personal/University/s2026/diploma/dataset/build_bcf/attendance_monitoring // ───────────────────────────────────── // Function: _addRecord // Address : 0x100000550 // ───────────────────────────────────── __int64 addRecord() { __int64 result; // x0 __int64 v1...
115
1,020
attendance_monitoring
original
build_fla_sub
fla+sub
#include <stdio.h> #include <stdlib.h> #include <string.h> // Define the maximum number of records and length of names #define MAX_RECORDS 100 #define MAX_NAME_LEN 50 // Define the structure to store attendance records struct AttendanceRecord { int id; char name[MAX_NAME_LEN]; int isPresent; }; // Declar...
// Decompiled with IDA Pro Hex-Rays // Input: /Volumes/Personal/University/s2026/diploma/dataset/build_fla_sub/attendance_monitoring // ───────────────────────────────────── // Function: _addRecord // Address : 0x100000550 // ───────────────────────────────────── __int64 addRecord() { int v0; // w8 int v1; // w8 ...
115
1,445
attendance_monitoring
original
build_fla_bcf
fla+bcf
"#include <stdio.h>\n#include <stdlib.h>\n#include <string.h>\n\n// Define the maximum number of rec(...TRUNCATED)
"// Decompiled with IDA Pro Hex-Rays\n// Input: /Volumes/Personal/University/s2026/diploma/dataset/b(...TRUNCATED)
115
7,293
attendance_monitoring
original
build_sub_bcf
sub+bcf
"#include <stdio.h>\n#include <stdlib.h>\n#include <string.h>\n\n// Define the maximum number of rec(...TRUNCATED)
"// Decompiled with IDA Pro Hex-Rays\n// Input: /Volumes/Personal/University/s2026/diploma/dataset/b(...TRUNCATED)
115
1,263
attendance_monitoring
original
build_all
fla+sub+bcf
"#include <stdio.h>\n#include <stdlib.h>\n#include <string.h>\n\n// Define the maximum number of rec(...TRUNCATED)
"// Decompiled with IDA Pro Hex-Rays\n// Input: /Volumes/Personal/University/s2026/diploma/dataset/b(...TRUNCATED)
115
8,695
bank
original
build_plain
"#include <stdio.h>\n#include <string.h>\n\nvoid create_account();\nvoid deposit_money();\nvoid with(...TRUNCATED)
"// Decompiled with IDA Pro Hex-Rays\n// Input: /Volumes/Personal/University/s2026/diploma/dataset/b(...TRUNCATED)
166
339
bank
original
build_sub
sub
"#include <stdio.h>\n#include <string.h>\n\nvoid create_account();\nvoid deposit_money();\nvoid with(...TRUNCATED)
"// Decompiled with IDA Pro Hex-Rays\n// Input: /Volumes/Personal/University/s2026/diploma/dataset/b(...TRUNCATED)
166
339
End of preview. Expand in Data Studio

OLLVM Obfuscation Impact on IDA Pro Hex-Rays Decompilation

Dataset Summary

A paired dataset for studying how compile-time OLLVM obfuscation (llvm-pass-hikari plugin, custom-built for LLVM 22) affects the quality of static decompilation (IDA Pro 9.2 Hex-Rays, ARM64).

Each row of the dataset is one decompilation: the full C source, the obfusCate-obfuscated version of that source, and the Hex-Rays pseudocode of the compiled binary, together with the variant metadata.

  • 800 rows = 50 programs × 2 source families × 8 OLLVM pass sets
  • 8 pass configurations per source family: none | sub | fla | bcf | fla+sub | fla+bcf | sub+bcf | fla+sub+bcf

Dataset Structure

Column Type Description
program string program name (50 unique)
source_family string original (compiled from plain C) or obfuscated (compiled from obfusCate output)
variant string build variant directory, e.g. build_fla_sub, build_obfuscated_all
passes string OLLVM pass set applied, e.g. fla+sub, fla+sub+bcf
source_code string the exact source file compiled for this variant: original C for source_family == "original", obfusCate v0.18.3 output for source_family == "obfuscated"
decompiled_code string full Hex-Rays pseudocode of the compiled binary
source_lines int64 line count of source_code
decompiled_lines int64 line count of decompiled_code

Raw files are also available in the repository:

data/train-00000-of-00001.parquet   800-row dataset (loads by default)
source_files/                       50 original C programs
obfuscated_source_files/            50 obfusCate-obfuscated versions
decompiled_pseudocode/build_*/      800 Hex-Rays pseudocode files (16 dirs × 50)

How to Use 

from datasets import load_dataset

ds = load_dataset("oleksiihrush/ollvm-decompilation-dataset")

row = ds["train"][0]
print(row["program"], "|", row["variant"], "|", row["passes"])
print(row["decompiled_code"])   # Hex-Rays pseudocode
print(row["source_code"])       # the exact source that was compiled for this variant

Every program appears in 16 rows: 8 with source_family == "original" (build_* variants) and 8 with source_family == "obfuscated" (build_obfuscated_* variants).

Pipeline 

source_files/*.c
    ├─ obfusCate v0.18.3 ──────────────► obfuscated_source_files/obfuscated_*.c
    └─ clang 22 (-O0) + OLLVM passes ──► 16 binary variants per program
                                              │
                            IDA Pro 9.2 Hex-Rays (ARM64)
                                              ▼
                            decompiled_pseudocode/build_*/   (800 .c files)
  • Compiler: Homebrew LLVM 22 clang, target arm64-apple-macos, -O0 (higher optimization levels would partially revert the obfuscation: jump threading unwinds flattening, instcombine unwinds substitution).
  • OLLVM plugin registered at PipelineStart + OptimizerLast; all modules enabled via per-module policy.json (e.g. bcf_prob: 50 per-function probability, sub_times: 1).
  • Obfuscation effectiveness verified at the binary level: the __TEXT machine-code section of every obfuscated binary differs from its plain counterpart.
  • Honest no-ops exist by design: sub has nothing to substitute in programs without integer arithmetic (e.g. crack1–crack8, bank, clock — string/strcmp-driven or float/struct-only code), and fla needs functions with ≥ 2 basic blocks (no-op for salary_slip).
  • Decompilation: ida -A -c -S export_pseudocode.py per binary, fresh analysis session per file; all 800 succeeded with zero failures. The // Input: header lines in decompiled files contain absolute paths of the original build environment and can be ignored.

Citation 

@dataset{hrushkovskyi2026ollvm,
  title     = {OLLVM Obfuscation Impact on IDA Pro Hex-Rays Decompilation},
  author    = {Hrushkovskyi, Oleksii},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/oleksiihrush/ollvm-decompilation-dataset},
  license   = {CC BY 4.0}
}

License

CC BY 4.0
Downloads last month
438
Number of rows:
800
Total file size:
143 MB