What causes "No valid examples after Whisper preprocessing" in unslothai/unsloth?

Check the per-example warnings logged during the split processing for the true exception.

unslothai/unsloth · error · ValueError

No valid examples after Whisper preprocessing

Error message

No valid examples after Whisper preprocessing

What it means

ValueError raised when the Whisper train split yields zero processed examples — every training row failed feature extraction/tokenization and was skipped during process_split(dataset, 'train'). Unlike the other audio paths this message has no skipped-count and only the train split is fatal; an empty eval split is tolerated. It means the per-example processing failed systemically, with details in the per-example warnings.

Source

Thrown at studio/backend/core/training/trainer.py:2503

                    logger.warning(f"Error processing Whisper {split_name} example {idx}: {e}")
                    skipped += 1
                    continue

                if (idx + 1) % 100 == 0:
                    self._update_progress(
                        status_message = f"Processing {split_name} audio... {idx + 1}/{len(ds)}"
                    )

            logger.info(
                f"Whisper {split_name} preprocessing: {len(processed)} examples ({skipped} skipped)\n"
            )
            return processed

        train_data = process_split(dataset, "train")
        eval_data = process_split(eval_dataset_raw, "eval") if eval_dataset_raw else None

        if not train_data:
            raise ValueError("No valid examples after Whisper preprocessing")

        return (train_data, eval_data)

    @staticmethod
    def _resolve_local_files(file_paths: list) -> list[str]:
        """Resolve a list of local dataset paths to concrete file paths."""
        all_files: list[str] = []
        for dataset_file in file_paths:
            if os.path.isabs(dataset_file):
                file_path = dataset_file
            elif os.path.exists(dataset_file):
                # A path relative to the current working directory (CLI usage)
                file_path = os.path.abspath(dataset_file)
            else:
                file_path = str(resolve_dataset_path(dataset_file))

            file_path_obj = Path(file_path)

View on GitHub (pinned to 203007d190)

Solutions

Check the per-example warnings logged during the split processing for the true exception.
Validate one row manually: dataset[0]['audio'] decodes and dataset[0]['text'] is a non-empty string.
Fix the column mapping — the most common cause is mapping a wrong/empty column to 'text'.
Drop or repair bad rows, then restart training.

Example fix

// before: mapped wrong column to text
mapping = {'audio': 'audio', 'file': 'text'}  # 'file' holds paths, not transcripts
// after
mapping = {'audio': 'audio', 'sentence': 'text'}

Defensive patterns

Strategy: validation

Validate before calling

def whisper_examples_valid(dataset, n=3) -> bool:
    for i in range(min(n, len(dataset))):
        a = dataset[i]['audio']
        t = dataset[i]['text']
        if not (a.get('array') is not None and len(a['array']) > 0 and isinstance(t, str) and t.strip()):
            return False
    return True

assert whisper_examples_valid(dataset), "Whisper preprocessing would skip all examples"

Try / catch

try:
    train_data, eval_data = trainer._preprocess_whisper_dataset(dataset, mapping, eval_split=True)
except ValueError as e:
    if 'No valid examples after Whisper preprocessing' in str(e):
        # check per-split warnings: usually a wrong text mapping or missing audio
        ...

Prevention

Confirm the mapped 'text' column really contains transcripts (not paths or labels).
Validate that audio decodes and text is non-empty for sample rows before training.
Distinguish train vs eval splits when debugging — only an empty train split raises.

When it happens

Trigger: All training rows fail in process_split: audio that cannot be decoded/resampled to 16kHz, empty or null transcripts producing tokenization errors, or audio feature extractor failures (e.g. zero-length clips).

Common situations: Dataset where the text column is actually empty/None for all rows after mapping the wrong column; audio bytes missing (paths/URLs unresolvable); corrupted audio encodings after a partial upload.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/1b7f059f65976098. Report an issue: GitHub.