metadata

license: apache-2.0

AReaL-boba-2-RL-Code Dataset

This dataset contains the training and testing data used for reinforcement learning in the AReaL-boba-2 model on coding tasks.

Dataset Structure

train/: Contains the training data for RL fine-tuning.
code_benchmark/: Contains the evaluation benchmark, organized into multiple subfolders, each corresponding to a specific coding benchmark suite, Codeforces, Code Contests and LiveCodeBench (v5) are supported now.

How to Use

To train and evaluate models using this dataset with AReaL:

Download the dataset from this page.
Train the model by following the AReaL training guideline.
Prepare the evaluation data:
- Move each subfolder inside code_benchmark/ to the evaluation/data/ directory in the AReaL repository.
Run evaluation according to the AReaL evaluation guideline.

Convert to Qwen3 prompt format

import json
with open("/path/to/raw.jsonl") as f, \
    open("/path/to/qwen3_data.jsonl", "w") as fout:

    for line in f:
        data = json.loads(line)

        prompt = data['prompt']
        prompt = prompt.replace(
            "<｜User｜>", "<|im_start|>user\n"
        ).replace(
            "<｜Assistant｜>", "\n/think<|im_end|>\n<|im_start|>assistant\n"
        )
        data['prompt'] = prompt

        print(json.dumps(data), file=fout)

License

Please refer to the LICENSE file associated with this dataset or contact the maintainers for more details.