README.md

metadata

task_categories:
  - translation
language:
  - en
tags:
  - llvm
  - gimple
  - gcc
  - clang
  - compiler
  - intermediate-representation
  - compiler-ir
  - ir
  - the-stack
pretty_name: TheStack-IRIS
size_categories:
  - 100K<n<1M

Dataset Card for TheStack-IRIS

TheStack-IRIS is a training dataset for GIMPLE IR to LLVM IR translation, derived from The Stack v1 (dedup). It contains ~310k C code samples paired with their corresponding GIMPLE and LLVM intermediate representations.

Dataset Structure

The dataset extends The Stack with additional IR fields:

• content: C source code
• irs: Dictionary containing intermediate representations:
  • gcc_gimple: GIMPLE IR generated by GCC v15.2.0
  • clang: LLVM IR generated by Clang v22.0.0
• max_stars_repo_name: Repository name for attribution
• max_stars_repo_licenses: Repository licenses

Usage

from datasets import load_dataset

# Load dataset
ds = load_dataset("HPAI-BSC/TheStack-IRIS", split="train")

# Example: access a sample
sample = ds[0]
c_code = sample["content"]
gimple_ir = sample["irs"]["gcc_gimple"]
llvm_ir = sample["irs"]["clang"]
repo_name = sample["max_stars_repo_name"]

Dataset Creation

TheStack-IRIS is created by filtering The Stack v1 (dedup) C samples and extending them with intermediate representations:

Stage	# Samples	Description
The Stack C samples	8,625,559	Original C code from The Stack
License filtering	8,136,034	Samples with permissive open-source licenses
Compile successfully	867,101	Samples that compile with both GCC v15.2.0 and Clang v22.0.0
IR generation	306,841	Samples with GIMPLE and LLVM IR aligned pairs

Filtering Process

1. Selected C language samples from The Stack v1 (dedup)
2. Compiled each sample with GCC v15.2.0 and Clang v22.0.0 at -O0
3. Retained samples that compile successfully with both compilers
4. Generated GIMPLE using gcc -fdump-tree-gimple -O0
5. Generated LLVM IR using clang -emit-llvm -O0

The dataset preserves all original The Stack metadata for proper attribution.

License

The dataset inherits licenses from The Stack. We used a subset of the top-20 permissive licenses. Each sample includes repository license information in the max_stars_repo_licenses field for proper attribution.

The licenses of this subset include:

- "MIT"
- "Apache-2.0"
- "BSD-3-Clause"
- "BSD-2-Clause"
- "CC0-1.0"
- "WTFPL"
- "0BSD"
- "FSFAP"
- "Zlib"
- "ECL-2.0"
- "NCSA"
- "BSD-Source-Code"
- "MIT-0"
- "BSL-1.0"
- "X11"
- "Intel"
- "ISC"
- "BSD-2-Clause-FreeBSD"
- "CNRI-Python"
- "BSD-3-Clause-Clear"
- "PSF-2.0"
- "LPL-1.02"

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}
}

And the original The Stack dataset:

@article{kocetkov2022stack,
  title={The Stack: 3 TB of permissively licensed source code},
  author={Kocetkov, Denis and Li, Raymond and Allal, Loubna Ben and Li, Jia and Mou, Chenghao and Ferrandis, Carlos Mu{\~n}oz and Jernite, Yacine and Mitchell, Margaret and Hughes, Sean and Wolf, Thomas and others},
  journal={arXiv preprint arXiv:2211.15533},
  year={2022}
}