karpathy/nanochat · error · ValueError

Unknown part type: {part['type']}

Error message

Unknown part type: {part['type']}

What it means

In `render_conversation`, an assistant message's content may be a list of typed parts. The supported part types are 'text', 'python' (a supervised tool call wrapped in <|python_start|>/<|python_end|>), and 'python_output' (unsupervised tool output wrapped in <|output_start|>/<|output_end|>). Any other value in part['type'] raises ValueError.

Source

Thrown at nanochat/tokenizer.py:216

                elif isinstance(content, list):
                    for part in content:
                        value_ids = self.encode(part["text"])
                        if part["type"] == "text":
                            # string part => simply add the tokens
                            add_tokens(value_ids, 1)
                        elif part["type"] == "python":
                            # python tool call => add the tokens inside <|python_start|> and <|python_end|>
                            add_tokens(python_start, 1)
                            add_tokens(value_ids, 1)
                            add_tokens(python_end, 1)
                        elif part["type"] == "python_output":
                            # python output => add the tokens inside <|output_start|> and <|output_end|>
                            # none of these tokens are supervised because the tokens come from Python at test time
                            add_tokens(output_start, 0)
                            add_tokens(value_ids, 0)
                            add_tokens(output_end, 0)
                        else:
                            raise ValueError(f"Unknown part type: {part['type']}")
                else:
                    raise ValueError(f"Unknown content type: {type(content)}")
                add_tokens(assistant_end, 1)

        # truncate to max_tokens tokens MAX (helps prevent OOMs)
        ids = ids[:max_tokens]
        mask = mask[:max_tokens]
        return ids, mask

    def visualize_tokenization(self, ids, mask, with_token_id=False):
        """Small helper function useful in debugging: visualize the tokenization of render_conversation"""
        RED = '\033[91m'
        GREEN = '\033[92m'
        RESET = '\033[0m'
        GRAY = '\033[90m'
        tokens = []
        for i, (token_id, mask_val) in enumerate(zip(ids, mask)):
            token_str = self.decode([token_id])

View on GitHub (pinned to 92d63d4e8b)

Solutions

1. Map every assistant part to type 'text', 'python', or 'python_output' when preparing data.
2. Convert OpenAI tool_calls entries: the call arguments become a 'python' part, the tool result becomes a 'python_output' part, plain prose becomes 'text'.
3. Write a small pre-flight check over your dataset (see validation code) before training.

Example fix

# before
{"role": "assistant", "content": [{"type": "tool_call", "text": "2+2"}]}

# after
{"role": "assistant", "content": [{"type": "python", "text": "2+2"}]}

VALID_PART_TYPES = {'text', 'python', 'python_output'}
for msg in conversation["messages"]:
    if isinstance(msg.get("content"), list):
        for part in msg["content"]:
            assert part.get("type") in VALID_PART_TYPES, f"bad part type {part.get('type')!r}"

def is_valid_part(part) -> bool:
    return isinstance(part, dict) and part.get('type') in {'text', 'python', 'python_output'} and isinstance(part.get('text'), str)

When it happens

Trigger: Feeding a chat SFT dataset where assistant parts use other type labels — e.g. OpenAI-style 'tool_calls'/'function_call', 'tool_response', 'code', or a typo like 'pythonoutput' — into tokenizer.render_conversation.

Common situations: Converting external chat datasets (ShareGPT/OpenAI format) into nanochat's conversation format without mapping tool-call fields; hand-writing synthetic conversations with invented part types; dataset schema drift after an upstream update.

Related errors

• Unknown content type: {type(content)}
• Invalid input type: {type(text)}
• Unknown dataset tag: {dataset_tag}
• Unsupported task type: {task_type}
• '{formula}': timed out after {duration} seconds

AI-assisted analysis of karpathy/nanochat@92d63d4e8b (2026-08-15). Data as JSON: /api/errors/a2abe98c17fc21ea. Report an issue: GitHub.