Please download images.tar to your local disk and use tar -xvf images.tar to unarchive the image files.

This dataset was mixed with hiyouga/geometry3k and hiyouga/math12k using the following script.

import os
from functools import partial

from datasets import DatasetDict, Features, Sequence, Value, concatenate_datasets, load_dataset


def process_sample(example: dict, index: int, split: str):
    if "images" in example:
        image = example["images"][0]
        image_path = os.path.join("images", split, f"{index}.png")
        image.save(image_path)
        images = [image_path]
    else:
        images = []

    return {
        "images": images,
        "problem": example["problem"],
        "answer": example["answer"],
    }


def main():
    geo3k = load_dataset("hiyouga/geometry3k")
    math12k = load_dataset("hiyouga/math12k")
    os.makedirs(os.path.join("images", "train"), exist_ok=True)
    os.makedirs(os.path.join("images", "test"), exist_ok=True)
    map_kwargs = {
        "with_indices": True,
        "num_proc": 64,
        "features": Features(
            {"images": Sequence(Value("string")), "problem": Value("string"), "answer": Value("string")}
        ),
    }
    geo3k_train = geo3k["train"].select(range(1500)).map(partial(process_sample, split="train"), **map_kwargs)
    geo3k_test = geo3k["test"].select(range(300)).map(partial(process_sample, split="test"), **map_kwargs)
    math12k_train = math12k["train"].select(range(1500)).map(partial(process_sample, split="train"), **map_kwargs)
    math12k_test = math12k["test"].select(range(300)).map(partial(process_sample, split="test"), **map_kwargs)
    trainset = concatenate_datasets([geo3k_train, math12k_train])
    testset = concatenate_datasets([geo3k_test, math12k_test])
    dataset = DatasetDict({"train": trainset, "test": testset})
    dataset.push_to_hub("hiyouga/rl-mixed-dataset")


if __name__ == "__main__":
    main()