SFT with tool-key failed / Without tool-key training example

Without training example with 'tool-key'

When the tool-key argument is provided, the prompt field gets converted to a string type after apply_chat_template processing, but downstream code expects it to remain as List[Dict]. This causes a type mismatch error.

prompt = data[prompt_key]
    if apply_chat_template:
        if tool_key is not None:
            tools = data[tool_key]
            if isinstance(tools, str):
                tools = json.loads(tools)
            elif isinstance(tools, np.ndarray):
                tools = tools.tolist()
            assert isinstance(tools, list), f"tools must be a list, got {type(tools)} instead"
        else:
            tools = None
        prompt = tokenizer.apply_chat_template(prompt, tools, tokenize=False, add_generation_prompt=True)

also the following messages is Str

for sample in samples:
    (sample,) = sample
    messages = sample.prompt
    token_ids, loss_mask = MASK_GENERATOR.get_loss_mask(messages)

But later code requires messages to be List[Dict]

def get_loss_mask(self, messages: List[Dict]) -> List[int]:
    if self.tokenizer_type == "qwen":
        if "<｜Assistant｜>" in self.tokenizer.get_added_vocab():
            return self.gen_multi_turn_loss_mask_distill_qwen(messages)

        return self.gen_multi_turn_loss_mask_qwen(messages)
    elif self.tokenizer_type == "distill_qwen":
        return self.gen_multi_turn_loss_mask_distill_qwen(messages)
    else:
        raise ValueError(f"Unsupported tokenizer type: {self.tokenizer_type}")

Start by tracing the tool_key handling around tokenizer.apply_chat_template and the later sample.prompt passed to MASK_GENERATOR.get_loss_mask. Reproduce SFT with a tool-key training example, then verify that the loss-mask path receives the expected List[Dict] messages without the reported type mismatch.