unslothai/unsloth · error · ValueError

{mode_title} training requires at least one string 'text' va

Error message

{mode_title} training requires at least one string 'text' value in {split_scope}; all {dropped_rows} rows were null or non-string.

What it means

Raised by raw-text dataset preparation after filtering: every row in the split had a null or non-string 'text' value, so the filtered dataset is empty and the requested training mode has nothing to train on. Partial drops only produce a warning notice; only the all-rows-dropped case raises. The message includes the mode title and split scope for context.

Source

Thrown at studio/backend/utils/datasets/raw_text.py:97

    # dropped rows or verify the result is non-empty without consuming the whole
    # stream. Keep the filter, skip only the len()-based diagnostics.
    if not hasattr(dataset, "__len__"):
        return filtered_dataset, [
            RawTextNotice(
                message = (
                    f"{mode_title}: streaming dataset — rows with null or "
                    f"non-string 'text' in {split_scope} are dropped on the fly."
                ),
                level = "info",
            )
        ]

    dropped_rows = len(dataset) - len(filtered_dataset)
    if not dropped_rows:
        return filtered_dataset, []

    if len(filtered_dataset) == 0:
        raise ValueError(
            f"{mode_title} training requires at least one string 'text' value "
            f"in {split_scope}; all {dropped_rows} rows were null or non-string."
        )

    return filtered_dataset, [
        RawTextNotice(
            message = (
                f"{mode_title}: dropped {dropped_rows:,} row(s) with null or "
                f"non-string 'text' values from {split_scope}"
            ),
            level = "warning",
            update_status = True,
        )
    ]


def prepare_raw_text_dataset(
    dataset: Dataset,

View on GitHub (pinned to 203007d190)

Solutions

Inspect the split: ds['text'] dtype and null count (e.g. sum of nulls) to confirm it is entirely empty/non-string.
Select a split that actually contains text (usually 'train').
If the text lives under another column, rename it to 'text' or rely on the auto-select path when no 'text' column exists at all.
Cast non-string columns (e.g. int labels) with .cast_column('text', Value('string')) only if the values genuinely are text mislabeled by schema.
Re-export the dataset from source with correct types.

Example fix

# before
result = prepare_raw_text(ds.select_columns(["text"]))  # 'text' all NULL -> ValueError

# after
result = prepare_raw_text(ds.rename_column("body", "text"))  # real text column

Defensive patterns

Strategy: validation

Validate before calling

def split_has_string_text(ds, text_column="text") -> bool:
    col = ds[text_column]
    non_null = [v for v in col if isinstance(v, str) and v]
    return len(non_null) > 0

# guard: assert split_has_string_text(ds[split]) before prepare_raw_text

Type guard

def has_usable_text_column(dataset, column="text") -> bool:
    feats = dataset.features
    return column in feats and len(dataset.filter(
        lambda r: isinstance(r[column], str) and r[column].strip(), load_from_cache_file=False
    )) > 0

Try / catch

try:
    filtered, notices = prepare_raw_text(ds, split_name=split)
except ValueError as e:
    if "requires at least one string" in str(e):
        ds = ds.rename_column(real_text_col, "text")
        filtered, notices = prepare_raw_text(ds, split_name=split)
    else:
        raise

Prevention

Check null counts and dtype of 'text' per split before training.
For streaming datasets rely on the on-the-fly drop notice instead of pre-filtering.
Rename the true text column to 'text' at load time.
Sanity-check that the chosen split actually contains data (len(ds[split]) > 0).

When it happens

Trigger: Calling raw text preparation on a dataset whose 'text' column is entirely null, or typed as non-string (e.g. all ints/floats/lists), for the selected split. For streaming datasets this error cannot fire (rows are dropped on the fly with an info notice instead).

Common situations: Dataset loaded with the wrong split that happens to have no text (e.g. a metadata-only aux split); text stored under a different column name while 'text' exists but is null; parquet column typed as binary/list after a bad conversion; user selected 'test' split of a dataset that only populates 'train'.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/4d9faa72408bf6d0. Report an issue: GitHub.