huggingface/open-r1 · error

Dataset Question Field Error: {prompt_column} is not support

Error message

Dataset Question Field Error: {prompt_column} is not supported.

What it means

make_conversation builds the GRPO 'prompt' column by reading example[prompt_column] (default 'prompt', configurable via dataset_prompt_column). If the mapped dataset lacks that key, it raises this ValueError, since the reward/policy pipeline requires a prompt extracted from a known field.

Source

Thrown at src/open_r1/grpo.py:98

    ##############
    # Load model #
    ##############
    logger.info("*** Loading model ***")
    model = get_model(model_args, training_args)

    # Get reward functions from the registry
    reward_funcs = get_reward_funcs(script_args)

    # Format into conversation
    def make_conversation(example, prompt_column: str = script_args.dataset_prompt_column):
        prompt = []

        if training_args.system_prompt is not None:
            prompt.append({"role": "system", "content": training_args.system_prompt})

        if prompt_column not in example:
            raise ValueError(f"Dataset Question Field Error: {prompt_column} is not supported.")

        prompt.append({"role": "user", "content": example[prompt_column]})
        return {"prompt": prompt}

    dataset = dataset.map(make_conversation)

    for split in dataset:
        if "messages" in dataset[split].column_names:
            dataset[split] = dataset[split].remove_columns("messages")

    #############################
    # Initialize the GRPO trainer
    #############################
    trainer = GRPOTrainer(
        model=model,
        reward_funcs=reward_funcs,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],

View on GitHub (pinned to 1416fa0cf2)

Solutions

1. Rename/prepare your dataset to expose the expected column: dataset = dataset.rename_column("question", "prompt").
2. Or set the matching CLI/config option (e.g. dataset_prompt_column="question") so prompt_column points at your field.
3. Print dataset.column_names to confirm the actual field names before training.

Example fix

// before
dataset = load_dataset("gsm8k", "main", split="train")  # has 'question', not 'prompt'
// after
dataset = load_dataset("gsm8k", "main", split="train").rename_column("question", "prompt")

required = training_args.dataset_prompt_column or "prompt"
if required not in dataset.column_names:
    raise SystemExit(f"Dataset lacks prompt column '{required}'; has {dataset.column_names}")

def has_prompt_column(example, col="prompt") -> bool:
    return isinstance(example, dict) and col in example and isinstance(example[col], str)

try:
    dataset = dataset.map(make_conversation)
except ValueError as e:
    if "Dataset Question Field Error" in str(e):
        sys.exit("Set the correct prompt column via the dataset prompt column config or rename the field")
    raise

When it happens

Trigger: Calling GRPOTrainer with a dataset whose prompt field is named something other than the configured prompt_column, e.g. a dataset with 'question' or 'problem' but no 'prompt', while training_args.system_prompt / prompt column config still defaults to 'prompt'.

Common situations: Swapping in a custom RL dataset (gsm8k, math) whose fields differ from the template's; forgetting to pass dataset_prompt_column; dataset.map over a split that lost the column after filtering/renaming.

Related errors

• Either `dataset_name` or `dataset_mixture` must be provided
• dataset_mixture must be a dictionary with a 'datasets' key.
• 'datasets' must be a list of dataset configurations
• Column names must be consistent across all dataset configura
• max_penalty {max_penalty} should not be positive

AI-assisted analysis of huggingface/open-r1@1416fa0cf2 (2026-08-30). Data as JSON: /api/errors/b217d6655df96384. Report an issue: GitHub.