LGL-DynT4: A Multi-Modal Benchmark with Aligned Dynamic Traces for Semantic Clone Detection

LGL-DynT4 is a multi-modal benchmark for semantic clone detection, with aligned source code, LLVM IR, and dynamic execution traces.

This Hugging Face repository is the benchmark/data release.
The corresponding development and reproducibility repository is maintained on GitHub:

GitHub:
https://github.com/lngooo/LGL-DynT4-Dev

Overview

LGL-DynT4 is designed for studying semantic similarity beyond surface-level lexical overlap. It provides paired static and dynamic views of programs under both Natural and Obfuscated settings.

The dataset includes:

• 50 functional seeds
• 21 implementation variants for each seed
• 10 execution traces for each implementation

Each sample is aligned across the following modalities:

• Source_Raw: original C source code
• Source_Clean: cleaned C source code
• LLVM_IR_Raw: raw LLVM IR
• LLVM_IR_Clean: cleaned LLVM IR
• Trace: dynamic execution traces

Versions

Natural

The original version with structurally diverse but semantically equivalent implementations.

Obfuscated

A version with identifier randomization applied to variable and function names, reducing lexical cues and making the task more dependent on structural and behavioral evidence.

Available Configurations

• manifest
• natural_source_raw
• natural_source_clean
• natural_ir_raw
• natural_ir_clean
• natural_trace
• obfus_source_raw
• obfus_source_clean
• obfus_ir_raw
• obfus_ir_clean
• obfus_trace

Loading the Dataset

from datasets import load_dataset

source_ds = load_dataset("lngoo/LGL-DynT4", "natural_source_clean")
trace_ds = load_dataset("lngoo/LGL-DynT4", "obfus_trace")
manifest_ds = load_dataset("lngoo/LGL-DynT4", "manifest")

Repository Structure

.
├── Data/
│   ├── Natural/
│   │   ├── Source_Raw/
│   │   ├── Source_Clean/
│   │   ├── LLVM_IR_Raw/
│   │   ├── LLVM_IR_Clean/
│   │   └── Trace/
│   └── Obfuscated/
│       ├── Source_Raw/
│       ├── Source_Clean/
│       ├── LLVM_IR_Raw/
│       ├── LLVM_IR_Clean/
│       └── Trace/
└── Manifest/

Citation

The paper is currently under review. Please replace the following placeholder information after acceptance.

@article{lgldynt4_under_review,
  title   = {LGL-DynT4: A Multi-Modal Benchmark with Aligned Dynamic Traces for Semantic Clone Detection},
  author  = {To be updated},
  journal = {Under Review},
  year    = {2026},
  note    = {Placeholder citation. Please replace with the final publication information.}
}