IRIS-14B

IRIS-14B is the first transformer model fine-tuned for neural translation between compiler Intermediate Representations (IRs) (IR-to-IR), specifically from GIMPLE (as emitted by GCC) to LLVM IR (as emitted by LLVM).

The model is initialized from Qwen/Qwen3-14B-Base (base model) and trained on aligned pairs of GIMPLE and LLVM IR extracted from real-world C programs.

Intended Use

IRIS-14B is designed for IR-to-IR translation, enabling interoperability between the GCC and LLVM compiler ecosystems by translating unoptimized GIMPLE into semantically equivalent LLVM IR that can be compiled by the LLVM toolchain.

Typical workflow:

1. 🛠️ Compile C code with GCC and dump the GIMPLE form using -fdump-tree-gimple.
2. 🔄 Use IRIS-14B to translate GIMPLE → LLVM IR.
3. ⚙️ Compile the generated LLVM IR using the LLVM toolchain (e.g., llc, clang).

Target Users:

• Compiler researchers and engineers.
• Toolchain developers.
• Advanced users working with GCC / LLVM internals.
• Researchers studying LLM-based program transformation.

Example Applications:

IRIS-14B enables compilation workflows that are otherwise unsupported, including:

• Legacy languages supported by GCC but not LLVM (e.g., Modula-2).
• GCC-specific C extensions not supported by Clang (e.g., nested functions).
• Fortran features supported in gfortran but not in flang.
• Ada programs relying on Scalar_Storage_Order, supported by GNAT (GCC-based) but not GNAT-LLVM.

In addition to source languages, IRIS facilitates cross-toolchain experimentation, including the reuse of optimizations and a decoupling of source-language support from target architecture availability.

Out-of-Scope Uses:

IRIS-14B is not intended for:

• Source-to-source translation (e.g., C → Rust).
• Optimized IR generation.
• General-purpose code generation or natural language tasks.

Quickstart

The following contains a code snippet illustrating how to use the model to generate LLVM IR based on a given GIMPLE sample.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "HPAI-BSC/IRIS-14B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input (GIMPLE sample)
sample = """
\nint main ()\n{\n  int D.2132;\n\n  {\n    long long int total;\n    int i;\n\n    scanf ("%d\\n", &n);\n    i = 0;\n    goto <D.2118>;\n    <D.2117>:\n    _1 = &A[i];\n    scanf ("%d", _1);\n    i = i + 1;\n    <D.2118>:\n    n.0_2 = n;\n    if (i < n.0_2) goto <D.2117>; else goto <D.2115>;\n    <D.2115>:\n    left[0] = 1;\n    _3 = A[0];\n    seen[_3] = 1;\n    i = 1;\n    goto <D.2122>;\n    <D.2121>:\n    _4 = i + -1;\n    _5 = left[_4];\n    left[i] = _5;\n    _6 = A[i];\n    _7 = seen[_6];\n    if (_7 == 0) goto <D.2128>; else goto <D.2129>;\n    <D.2128>:\n    _8 = A[i];\n    seen[_8] = 1;\n    _9 = left[i];\n    _10 = _9 + 1;\n    left[i] = _10;\n    <D.2129>:\n    i = i + 1;\n    <D.2122>:\n    n.1_11 = n;\n    if (i < n.1_11) goto <D.2121>; else goto <D.2119>;\n    <D.2119>:\n    memset (&seen, 0, 400004);\n    total = 0;\n    n.2_12 = n;\n    i = n.2_12 + -1;\n    goto <D.2126>;\n    <D.2125>:\n    _13 = A[i];\n    _14 = seen[_13];\n    if (_14 == 0) goto <D.2130>; else goto <D.2131>;\n    <D.2130>:\n    _15 = A[i];\n    seen[_15] = 1;\n    _16 = i + -1;\n    _17 = left[_16];\n    _18 = (long long int) _17;\n    total = total + _18;\n    <D.2131>:\n    i = i + -1;\n    <D.2126>:\n    if (i > 0) goto <D.2125>; else goto <D.2123>;\n    <D.2123>:\n    printf ("%lld\\n", total);\n    D.2132 = 0;\n    return D.2132;\n  }\n  D.2132 = 0;\n  return D.2132;\n}\n\n
"""

messages = [
    {"role": "user", "content": sample}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
print("prompt: ", text)

# model inference
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 
content = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")
print("model response (LLVM IR): ", content)

IRIS-14B inherits the 32,768-token native context length from its base model.

Limitations and bias

Language-dependent IR features

• IRIS-14B is trained only on C-derived IR.
• Therefore, it may not fully capture lowering patterns specific to other source languages (e.g., Ada, Fortran).

Translation direction

• Trained and evaluated on the GIMPLE → LLVM IR translation direction only.
• Reverse translation is not supported due to limited GCC support for parsing textual GIMPLE.

Training data

IRIS-14B is trained exclusively on IR pairs extracted from C-language programs, compiled with both GCC and Clang to obtain aligned IR pairs. To train the model, we construct two datasets:

• TheStack-IRIS: Derived from TheStack v1 dataset, filtered to deduplicated and compilable C code. Approximately 310k IR samples are retained after compilation filtering.

• GNU-IRIS: A dataset introduced with IRIS, consisting of around 13k aligned function-level IR pairs extracted from selected GNU utilities repositories.

Training details

• Base model: Qwen3 (14B parameters)
• Architecture: Decoder-only Transformer
• Training data: ~1.4B tokens of paired IRs
• Max sequence length: 16,384 tokens
• Epochs: 3
• Optimizer: AdamW (β₁ = 0.9, β₂ = 0.999)
• Learning rate: peak at 2.07e-5
• Scheduler: cosine with 3% warmup
• Hardware: 15 nodes × 4 NVIDIA H100 GPUs
• Training time: ~35 hours
• Energy consumption: ~0.86 MWh
• Estimated CO₂ emissions 🌱: ~244 kg

Evaluation results

IRIS-14B is evaluated on an IR-to-IR translation task, which consists of translating GIMPLE IR into LLVM IR such that the generated code successfully compiles with the LLVM toolchain and passes functional I/O tests. For evaluation, we construct two benchmarks:

• CodeForces-IRIS: IR snippets extracted from competitive programming code from CodeForces with higher structural complexity.
• ExeBench-IRIS: IR snippets extracted from the Exebench benchmarks consisting of simplified real-world code extracted from repositories.

For evaluation, we use two metrics reported as pass@1, N=3:

• Compile: Percentage of samples where the generated LLVM IR compiles successfully (pass@1, N=3)
• I/O Tests: Percentage of samples that compile and pass all functional tests (pass@1, N=3)

🏆 Leaderboard (pass@1, N=3)

Last leaderboard update: Sept 2026.

The IR-to-IR translation task is challenging for general-purpose and coder LLMs, and model size alone correlates weakly with performance. Despite having only 14B parameters, IRIS-14B consistently outperforms all larger baseline models.

📖 Citation

If you use IRIS-14B, please cite our work:

@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}
}