To ensure robust defensive programming and prevent cryptic runtime errors (such as TypeError during string concatenation or iteration), we should explicitly validate that mask_out_tokens is either a string or a sequence of strings.

    mask_out_sequences = []
    if mask_out_tokens:
        if isinstance(mask_out_tokens, str):
            mask_out_tokens = [mask_out_tokens]
        elif not isinstance(mask_out_tokens, (list, tuple)):
            raise TypeError("Unsloth: mask_out_tokens must be a string, list, or tuple of strings.")
        for token_string in mask_out_tokens:
            if not isinstance(token_string, str):
                raise TypeError(f"Unsloth: mask_out_tokens elements must be strings, but got {type(token_string).__name__}")
            for candidate in dict.fromkeys((token_string, " " + token_string)):
                ids = tokenizer(candidate, add_special_tokens = False).input_ids
                if ids and ids not in mask_out_sequences:
                    mask_out_sequences.append(ids)
    pass

Propagate mask_out_tokens to vision collators

When callers use this new kwarg with a vision trainer, train_on_responses_only takes the later _is_vision_collator branch and assigns data_collator.train_on_responses_only from a recursive call at lines 695-702, but that call omits mask_out_tokens, so the inner closure builds an empty mask_out_sequences. This means VLM response-only training still includes </think> or any requested token in the loss despite accepting mask_out_tokens=[...]; pass the kwarg through the collator masking setup as well.

Useful? React with 👍 / 👎.

Fixed in 76587db — the vision-collator recursive call now passes mask_out_tokens through, so the VLM path applies the same re-masking instead of silently ignoring the kwarg. Text-path label output is regression-identical.

Forward mask_out_tokens through the MLX wrapper

When using the MLX API (unsloth_zoo.mlx.trainer.train_on_responses_only), this newly added parameter cannot be used because the wrapper signature at unsloth_zoo/mlx/trainer.py:2877-2886 and its call into this helper at 2933-2941 were not updated. Passing mask_out_tokens=[...] through the documented MLX mirror of the HF/Unsloth API raises TypeError before it reaches this implementation, so MLX training cannot apply the new masking behavior.

Useful? React with 👍 / 👎.

Fixed in 76587db — the MLX wrapper signature now accepts mask_out_tokens and forwards it to the HF implementation, mirroring the rest of the API surface.