Dataset Card for GNU-IRIS

GNU-IRIS is a training dataset for GIMPLE IR to LLVM IR translation, derived from GNU utilities source code. It contains 13,049 C functions paired with their corresponding GIMPLE and LLVM intermediate representations.

Dataset Structure

The default dataset combines GNU utils projects into a single training dataset. However, you can also access each GNU utility individually as a split:

Configuration	Package	# Samples
default	All utilities	13,049
mailutils	GNU Mailutils	8,916
coreutils	GNU Coreutils	1,307
recutils	GNU Recutils	1,118
inetutils	GNU Inetutils	1,026
findutils	GNU Findutils	482
diffutils	GNU Diffutils	200

Data Fields

Each sample contains the following fields:

• filename: Path to the file in the original project.
• function: Function name inside the given file.
• gimple: GIMPLE IR representation generated by GCC v15.
• llvm: LLVM IR representation generated by Clang v22.

Usage

from datasets import load_dataset

# Load the complete dataset (default config)
ds = load_dataset("HPAI-BSC/GNU-IRIS", split="train")

# Or load a specific utility
ds_coreutils = load_dataset("HPAI-BSC/GNU-IRIS", "coreutils", split="train")

# Example: access a sample
sample = ds[0]
instruction = sample["instruction"]
gimple_ir = sample["input"]
llvm_ir = sample["output"]

Dataset Creation

GNU-IRIS is created by extracting C functions from GNU utilities source code and generating their intermediate representations using GCC and Clang compilers through their Makefiles.

1. Compile each GNU utils project with GCC v15 and Clang v22.
2. Generate GIMPLE using gcc -fdump-tree-gimple -O0
3. Generate LLVM IR using clang -emit-llvm -O0
4. Parse GIMPLE and LLVM IR files to extract pairs of equivalent functions

License

The dataset inherits licenses from each GNU project (generally GPL V3+).

Citation

If you use this dataset, please cite:

@article{ramirez2026llm,
  title={LLM Translation of Compiler Intermediate Representation},
  author={Valenzuela-Ramirez, Andrea and Gutierrez-Gomez, Cristian and Barroso, Marta and Garcia-Gasulla, Dario and Royuela, Sara},
  journal={arXiv preprint arXiv:2605.08247},
  year={2026}
}