AReaL: Skill for Claude Code

Add Dataset

Add a new dataset loader to AReaL.

When to Use

This skill is triggered when:

• User asks "how do I add a dataset?"
• User wants to integrate a new dataset
• User mentions creating a dataset loader

Step-by-Step Guide

Step 1: Create Dataset File

Create areal/dataset/<name>.py:

from datasets import Dataset, load_dataset


def get_<name>_sft_dataset(
    path: str,
    split: str,
    tokenizer,
    max_length: int | None = None,
) -> Dataset:
    """Load dataset for SFT training.

    Args:
        path: Path to dataset (HuggingFace hub or local path)
        split: Dataset split (train/validation/test)
        tokenizer: Tokenizer for processing
        max_length: Maximum sequence length (optional)

    Returns:
        HuggingFace Dataset with processed samples
    """
    dataset = load_dataset(path=path, split=split)

    def process(sample):
        # Tokenize the full sequence (prompt + response)
        seq_token = tokenizer.encode(
            sample["question"] + sample["answer"] + tokenizer.eos_token
        )
        prompt_token = tokenizer.encode(sample["question"])
        # Loss mask: 0 for prompt, 1 for response
        loss_mask = [0] * len(prompt_token) + [1] * (len(seq_token) - len(prompt_token))
        return {"input_ids": seq_token, "loss_mask": loss_mask}

    dataset = dataset.map(process).remove_columns(["question", "answer"])

    if max_length is not None:
        dataset = dataset.filter(lambda x: len(x["input_ids"]) <= max_length)

    return dataset


def get_<name>_rl_dataset(
    path: str,
    split: str,
    tokenizer,
    max_length: int | None = None,
) -> Dataset:
    """Load dataset for RL training.

    Args:
        path: Path to dataset
        split: Dataset split
        tokenizer: Tokenizer for length filtering
        max_length: Maximum sequence length

    Returns:
        HuggingFace Dataset with prompts and answers for reward computation
    """
    dataset = load_dataset(path=path, split=split)

    def process(sample):
        messages = [
            {
                "role": "user",
                "content": sample["question"],
            }
        ]
        return {"messages": messages, "answer": sample["answer"]}

    dataset = dataset.map(process).remove_columns(["question"])

    if max_length is not None:

        def filter_length(sample):
            content = sample["messages"][0]["content"]
            tokens = tokenizer.encode(content)
            return len(tokens) <= max_length

        dataset = dataset.filter(filter_length)

    return dataset