BootTrans

Official reproduction guide for BootTrans: Bootstrapping Code Translation with Weighted Multilanguage Exploration (ACL 2026).

BootTrans studies reinforcement-learning-based code translation. This repository contains the training pipeline, code-execution reward, validation data, and evaluation script used to reproduce the main experimental workflow.

🧰 1. Environment Setup

1.1 Create Python Environment

We recommend using a clean Python environment. The project is tested with Python 3.9+.

conda create -n boottrans python=3.12 -y
conda activate boottrans

pip install -r requirements.txt
pip install -e .

The training code depends on torch, transformers, vllm, ray, and other packages listed in requirements.txt. Please install the CUDA-compatible versions of PyTorch and vLLM according to your local GPU driver and CUDA version if the default installation is not suitable.

1.2 Install MultiPL-E Evaluation Package

BootTrans uses the MultiPL-E execution framework for code evaluation.

cd MultiPLE/evaluation
pip install -e .
cd ../..

1.3 Prepare Java Dependencies

Some Java evaluation cases require the bundled JavaFX environment.

unzip ./envs/javafx.zip -d ./envs/

The repository also provides envs/javatuples-1.2.jar for Java-related execution.

📦 2. Data

The released data files are placed under data/.

Split	File	Description
Training	data/kodcode.jsonl	Training data for BootTrans.
Validation	data/validation_humanevalx.jsonl	HumanEval-X validation set.
Validation	data/validation_transcoder.jsonl	TransCoder validation set.

Each JSONL sample follows the fields consumed by the training and evaluation scripts, including source prompt, solution, source language, and target language.

🏋️ 3. Training

The default training entry point is:

bash ./examples/boottrans.sh

By default, the script uses Qwen/Qwen3-1.7B as the base model:

MODEL_PATH=Qwen/Qwen3-1.7B

To reproduce experiments with another base model, modify both:

1. MODEL_PATH in examples/boottrans.sh
2. worker.actor.model.model_path in examples/qwen_lang.yaml

The main configuration file is examples/qwen_lang.yaml. Important fields include:

Field	Meaning
data.train_files	Training JSONL file.
data.val_files	Validation JSONL file used during training.
algorithm.adv_estimator	Advantage estimator; BootTrans uses grpo_lang.
worker.rollout.n	Number of sampled responses per prompt.
worker.rollout.tensor_parallel_size	Tensor parallel size for rollout.
trainer.n_gpus_per_node	Number of GPUs used on each node.
trainer.total_epochs	Number of training epochs.
trainer.save_freq	Checkpoint saving frequency.

For a single-node run, set trainer.n_gpus_per_node and worker.rollout.tensor_parallel_size according to the available GPUs. The default configuration uses 2 GPUs.

📊 4. Evaluation

Evaluation is implemented in evaluation/autochatmodel.py. The script loads a model with vLLM, generates translations for both validation sets, executes the generated programs, and reports execution accuracy by language pair.

⚙️ 4.1 Configure the Model

Edit the models dictionary in evaluation/autochatmodel.py:

models = {
    "qwen_base": "Qwen/Qwen3-1.7B",
}

The key is the output name, and the value is the Hugging Face model ID or a local checkpoint path.

▶️ 4.2 Run Evaluation

Run the evaluation script from the repository root:

python ./evaluation/autochatmodel.py

The script evaluates:

data/validation_transcoder.jsonl
data/validation_humanevalx.jsonl

and writes JSONL results to:

result/<model_name>_transcoder.jsonl
result/<model_name>_humanevalx.jsonl

📝 5. Notes

• Code execution evaluation requires the corresponding language runtimes to be available in the environment.
• If using a local checkpoint saved by training, set the checkpoint path in evaluation/autochatmodel.py.

📚 Citation

If you use this repository, please cite:

@inproceedings{boottrans2026,
  title = {Bootstrapping Code Translation with Weighted Multilanguage Exploration},
  booktitle = {ACL},
  year = {2026}
}