unslothai/unsloth · error · ValueError

{mode_title} training requires a string 'text' column but no

Error message

{mode_title} training requires a string 'text' column but none was found in {split_scope} (columns: {col_names}).

What it means

Raised by raw-text preparation when the dataset has no 'text' column AND no string-typed columns exist at all to auto-select. When string columns do exist, the first is auto-selected as training text (with a notice, and a disambiguation notice if there are several); this error is the terminal case where no candidate column exists.

Source

Thrown at studio/backend/utils/datasets/raw_text.py:130


def prepare_raw_text_dataset(
    dataset: Dataset,
    *,
    mode_label: str = "raw text",
    split_name: str | None = None,
    eos_token: str | None = None,
    append_eos: bool = False,
) -> RawTextPreparationResult:
    notices: list[RawTextNotice] = []
    mode_title = mode_label.capitalize()
    split_scope = _split_scope(split_name)

    col_names = resolve_column_names(dataset)
    if "text" not in col_names:
        string_cols = _string_columns(dataset)
        if not string_cols:
            raise ValueError(
                f"{mode_title} training requires a string 'text' column but none "
                f"was found in {split_scope} (columns: {col_names})."
            )

        renamed_col = string_cols[0]
        if len(string_cols) > 1:
            notices.append(
                RawTextNotice(
                    message = (
                        f"{mode_title}: dataset has {len(string_cols)} string "
                        f"columns ({string_cols}); auto-selecting '{renamed_col}' "
                        "as the training text. Rename the intended column to "
                        "'text' to override."
                    ),
                    level = "warning",
                    update_status = True,
                )
            )

View on GitHub (pinned to 203007d190)

Solutions

Check resolve_column_names(dataset) output (shown in the error) to see what columns actually exist.
If the dataset is pre-tokenized, use the tokenized/TRL training path instead of raw-text mode.
Load the original untokénized dataset revision and train from its string text column.
If a numeric column genuinely encodes text (e.g. byte values), convert it back to strings and rename to 'text'.
Pass a different split that contains the raw text column.

Example fix

# before
prepare_raw_text(load_dataset("user/pretokenized"))  # columns: input_ids, labels -> ValueError

# after
prepare_raw_text(load_dataset("user/raw-corpus"))  # columns: text -> ok

Defensive patterns

Strategy: type-guard

Validate before calling

def dataset_supports_raw_text(dataset) -> bool:
    """True when a 'text' column or at least one string column exists."""
    feats = dataset.features
    if "text" in feats:
        return True
    return any(
        getattr(f, "dtype", None) == "string" or isinstance(f, str)
        for f in feats.values()
    )

Type guard

def has_string_column(features: dict) -> bool:
    return any(getattr(f, "dtype", None) == "string" for f in features.values())

Try / catch

try:
    result = prepare_raw_text(ds)
except ValueError as e:
    if "requires a string 'text' column" in str(e):
        raise SystemExit(
            f"Dataset columns {list(ds.features)} contain no text; "
            "load the untokénized source dataset."
        ) from e
    raise

Prevention

Inspect dataset.features before selecting raw-text training mode.
Pre-tokenized datasets (input_ids/labels) belong on the tokenized training path.
Watch for the auto-select notice naming which string column was chosen — confirm it is the intended text.
When multiple string columns exist, rename the intended one to 'text' to avoid ambiguity.

When it happens

Trigger: Calling raw-text preparation on a dataset whose columns are all non-string (numeric features, label ids, images, lists), e.g. a pre-tokenized dataset with input_ids/attention_mask, or a pure tabular feature dataset.

Common situations: Loading an already-tokenized dataset (input_ids only) into a raw-text trainer; selecting a feature/embedding table dataset; datasets where text was replaced by numeric encodings during preprocessing.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/b6b14c8b694cc752. Report an issue: GitHub.