unslothai/unsloth · error · ValueError

Unknown or unsupported format: '{fmt}'

Error message

Unknown or unsupported format: '{fmt}'

What it means

ValueError from scan_dataset when the requested format (after FORMAT_ALIASES normalization) has no registered scanner. The format registry maps names like chatml/sharegpt/gptoss/alpaca to find_none_* scanners; auto-detection never raises this (unknown data returns clean stats with format='unknown'), so it always means an explicit fmt string was mis-typed or unsupported.

Source

Thrown at studio/backend/utils/datasets/dataset_none_detect.py:565

    conv_info = _probe_conversation(dataset)
    if was_auto:
        fmt = "unknown"
        for entry in FORMAT_REGISTRY:
            if entry["match"](dataset, conv_info):
                fmt = entry["name"]
                break
        # No format matched: return clean stats (format="unknown") instead of
        # raising, so callers can branch on stats["format"].
        if fmt == "unknown":
            return {
                "format": "unknown",
                "total_rows": len(dataset),
                "findings": [],
                "bad_row_indices": [],
            }
    scanner = get_scanner(fmt)
    if scanner is None:
        raise ValueError(f"Unknown or unsupported format: '{fmt}'")
    # Column forwarding: on auto-detect pass the probed column (the best
    # choice). On an explicit format let that scanner pick its own column, so
    # e.g. fmt='sharegpt' always scans 'conversations', not 'messages' (P1 fix);
    # gptoss has its own messages-first rule. alpaca never takes a column.
    use_probed_col = conv_info is not None and fmt != "alpaca" and was_auto
    if use_probed_col:
        stats = scanner(dataset, col = conv_info["column"])
    else:
        stats = scanner(dataset)
    stats["format"] = fmt
    return stats


# ---------------------------------------------------------------------------
# Report printing
# ---------------------------------------------------------------------------

View on GitHub (pinned to 203007d190)

Solutions

1. Use a supported format name: 'auto', or one of the registered scanners (chatml, sharegpt, gptoss, alpaca — check FORMAT_ALIASES/get_scanner for the exact set in your version)
2. Prefer fmt='auto' unless you specifically need to force a scanner — auto-detect never raises this error
3. Validate fmt against the registry at the CLI/API boundary: get_scanner(fmt) is None -> reject early

Example fix

# before
stats = scan_dataset(ds, fmt='chatml-v2')  # unsupported name

# after
stats = scan_dataset(ds, fmt='auto')
# or explicit supported name:
stats = scan_dataset(ds, fmt='chatml')

from studio.backend.utils.datasets.dataset_none_detect import get_scanner, FORMAT_ALIASES

def is_supported_fmt(fmt: str) -> bool:
    return get_scanner(FORMAT_ALIASES.get(fmt, fmt)) is not None

When it happens

Trigger: scan_dataset(ds, fmt='chatm') (typo), fmt='dialogue' (invented name), or an alias missing from FORMAT_ALIASES in this version. Note fmt='auto'/'unknown' bypass the raise.

Common situations: User-supplied format strings from CLI flags or UI dropdowns forwarded without validation; version skew where a format alias was added/renamed; mixing up trainer format names with scanner format names.

Related errors

• An audio sample should have one of 'path' or 'bytes' but bot
• No conversation column found. Expected one of {CONVERSATION_
• 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

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/d9bcb21337f2bea1. Report an issue: GitHub.