unslothai/unsloth · error · ValueError

Unknown completion masking template: {dataset_template}

Error message

Unknown completion masking template: {dataset_template}

What it means

ValueError from apply_completion_masking when an explicit dataset_template is passed but is not a key in TEMPLATE_TO_RESPONSES_MAPPER (the manual marker table in model_mappings). The explicit template bypasses tokenizer marker detection, so an unknown name cannot be silently ignored — masking would target the wrong tokens, so it raises instead. Note the lookup is truthiness-based, so an empty-string template also lands here.

Source

Thrown at studio/backend/utils/datasets/completion_masking.py:83

    the run instead of silently changing the training objective.
    """
    if notify is None:
        notify = lambda level, message: None
    kwargs = {}
    if num_proc is not None:
        kwargs["num_proc"] = num_proc

    processor = getattr(trainer, "processing_class", None) or getattr(trainer, "tokenizer", None)
    if type(processor).__name__ == "TokenizerWrapper":
        wrapped = getattr(processor, "_tokenizer", None)
        if wrapped is not None:
            processor = wrapped
    inner = getattr(processor, "tokenizer", processor)

    if dataset_template is not None:
        markers = TEMPLATE_TO_RESPONSES_MAPPER.get(dataset_template)
        if not markers:
            raise ValueError(f"Unknown completion masking template: {dataset_template}")
        has_preset_markers = hasattr(inner, "_unsloth_input_part") and hasattr(
            inner, "_unsloth_output_part"
        )
        if has_preset_markers:
            previous_instruction = inner._unsloth_input_part
            previous_response = inner._unsloth_output_part
            inner._unsloth_input_part = markers["instruction"]
            inner._unsloth_output_part = markers["response"]
            try:
                trainer = train_fn(trainer, **kwargs)
            finally:
                inner._unsloth_input_part = previous_instruction
                inner._unsloth_output_part = previous_response
        else:
            trainer = train_fn(
                trainer,
                instruction_part = markers["instruction"],
                response_part = markers["response"],

View on GitHub (pinned to 203007d190)

Solutions

1. Use a template name that exists as a key in model_mappings.TEMPLATE_TO_RESPONSES_MAPPER — inspect its keys to get the exact spelling
2. Pass dataset_template=None to fall back to tokenizer marker auto-detection with the manual model-table fallback
3. Validate the value against TEMPLATE_TO_RESPONSES_MAPPER at the API/config boundary before it reaches training

Example fix

# before
apply_completion_masking(trainer, model, train_fn, dataset_template='qwen-chat')  # typo

# after
from .model_mappings import TEMPLATE_TO_RESPONSES_MAPPER
if dataset_template is not None and dataset_template not in TEMPLATE_TO_RESPONSES_MAPPER:
    raise ValueError(f"template must be one of {sorted(TEMPLATE_TO_RESPONSES_MAPPER)}")
apply_completion_masking(trainer, model, train_fn, dataset_template=dataset_template)

from studio.backend.utils.datasets.model_mappings import TEMPLATE_TO_RESPONSES_MAPPER

def is_known_template(name: str | None) -> bool:
    return name is None or name in TEMPLATE_TO_RESPONSES_MAPPER

def known_templates() -> list[str]:
    return sorted(TEMPLATE_TO_RESPONSES_MAPPER)

When it happens

Trigger: Calling apply_completion_masking(..., dataset_template='chatml-qwen2') when the table only knows names like those registered in model_mappings.TEMPLATE_TO_RESPONSES_MAPPER; also dataset_template='' (falsy markers).

Common situations: Typos in template names from UI dropdowns or config files; version skew where the table was renamed between releases; passing a model name instead of a template name; forwarding a user free-text field without validation.

Related errors

• gradient_accumulation_steps must be >= 1
• lora_rank must be >= 1
• lora_alpha must be >= 1 (a zero/negative alpha scales the ad
• resolution must be a multiple of 8 and >= 64
• mixed_precision must be one of bf16 / fp16 / no

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/1dc03af6fe96d9ad. Report an issue: GitHub.