unslothai/unsloth · error · ValueError

No conversation column found. Expected one of {CONVERSATION_

Error message

No conversation column found. Expected one of {CONVERSATION_COLUMNS}, got columns: {dataset.column_names}

What it means

ValueError from find_none_chatml (the core None-turn scanner) when no conversation column could be determined: auto-probe (_probe_conversation over CONVERSATION_COLUMNS) found nothing, or the explicitly passed col is not among dataset.column_names. The scanner needs a column of conversation lists to walk turn-by-turn, so there is nothing to scan without one.

Source

Thrown at studio/backend/utils/datasets/dataset_none_detect.py:251

# ---------------------------------------------------------------------------


def find_none_chatml(dataset: Dataset, col: str = None) -> dict:
    """
    Scan chatml/sharegpt/gptoss dataset for turns with None/empty content.
    Auto-detects the conversation column if col=None.

    Returns a stats dict with a complete 'findings' list - one entry per bad
    turn with row_index, turn_index, role, value_type, and raw_value.
    """
    if col is None:
        # Reuse _probe_conversation so the all_corrupt path is handled here too.
        _cinfo = _probe_conversation(dataset)
        if _cinfo is not None:
            col = _cinfo["column"]

    if col is None or col not in dataset.column_names:
        raise ValueError(
            f"No conversation column found. "
            f"Expected one of {CONVERSATION_COLUMNS}, got columns: {dataset.column_names}"
        )

    stats = {
        "total_rows": len(dataset),
        "column": col,
        "rows_with_none_turns": 0,
        "total_none_turns": 0,
        "none_by_role": {},  # role -> count of None turns
        "none_by_type": {},  # "None" | "empty_string" | "whitespace_only" -> count
        "rows_all_none": 0,  # rows where every turn is bad
        "bad_row_indices": [],  # every row index that has at least one bad turn
        "findings": [],  # detailed per-turn list
    }

    for i, row in enumerate(dataset):
        conversation = row[col]

View on GitHub (pinned to 203007d190)

Solutions

1. Pass the actual column name explicitly: find_none_chatml(dataset, col='history')
2. Rename the column to a recognized name: dataset.rename_column('history', 'conversations')
3. If the dataset is Alpaca-style (no conversation column), use the alpaca scanner / fmt='alpaca' instead of the chatml one

Example fix

# before
stats = find_none_chatml(dataset)  # dataset has column 'history'

# after
stats = find_none_chatml(dataset, col='history')
# or: dataset = dataset.rename_column('history', 'conversations')

from studio.backend.utils.datasets.dataset_none_detect import CONVERSATION_COLUMNS

def has_conversation_column(dataset) -> bool:
    return any(c in dataset.column_names for c in CONVERSATION_COLUMNS)

def find_chat_column(dataset) -> str | None:
    cols = set(dataset.column_names)
    for c in CONVERSATION_COLUMNS:
        if c in cols:
            return c
    return None

When it happens

Trigger: Calling find_none_chatml(dataset) on a dataset whose columns lack any of CONVERSATION_COLUMNS (e.g. only 'instruction'/'output' Alpaca-style columns), or find_none_chatml(dataset, col='chat') when the column is actually named 'messages'.

Common situations: Running the None/blank-turn audit tool on the wrong format (Alpaca instead of chat), column renamed during a previous preprocessing step, schemas that embed conversations under a non-standard name like 'history' or 'dialog'.

Related errors

• An audio sample should have one of 'path' or 'bytes' but bot
• No valid conversation column found in {dataset.column_names}
• No valid conversation column found in {dataset.column_names}
• scan_dataset requires a single Dataset split, not a DatasetD
• Unknown or unsupported format: '{fmt}'

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/24a32fe3b240ee32. Report an issue: GitHub.