google

/

translategemma-4b-it

like 779

Follow

Google 56.2k

Hi All,

I am using this model for a translation task and observed some behavior related to the chat template and source_lang_code handling.

When the source language is known, I can specify it explicitly, and the model works correctly.

Example:

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "source_lang_code": "hi",   # Source language (Hindi)
                "target_lang_code": "en",   # Target language (English)
                "text": "नमस्ते, आप कैसे हैं?",
            }
        ],
    }
]

However, the issue occurs when the source language is unknown.

Case 1: Removing source_lang_code or using "auto"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "source_lang_code": "auto",   # or remove this line
                "target_lang_code": "en",
                "text": "नमस्ते, आप कैसे हैं?",
            }
        ],
    }
]

In this case, I receive a dictionary/key-related error.

Case 2: Setting the source language to "en" by default

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "source_lang_code": "en",
                "target_lang_code": "en",
                "text": "नमस्ते, आप कैसे हैं?",
            }
        ],
    }
]

Interestingly, the model still generates a proper response even though the actual source language is Hindi.

My Observation

It seems that:

• the model may already be capable of automatically detecting the source language internally,
• but the template currently requires source_lang_code to be explicitly provided,
• and "auto" is not yet supported.

Suggestion

It would be very useful if the template supported:

"source_lang_code": "auto"

or handled missing source_lang_code gracefully by automatically detecting the language.

This would improve usability for real-world multilingual translation pipelines where the source language may not always be known beforehand.

Please let me know if this behavior is expected, a bug, or if support for "auto" is planned in future updates.

Thank you.