ai4bharat

/

indictrans2-indic-en-1B

like 33

Follow

AI4Bharat 1.47k

Hi AI4Bharat Team,

I'm using the ai4bharat/indictrans2-en-indic-1B model for a thesis project. Until recently, everything worked perfectly, both in local and Colab environments. However, after a recent update, the model is throwing the following error during generate():

AttributeError: 'IndicTransConfig' object has no attribute 'vocab_size'

This appears to be because vocab_size is missing in the current config.json, and Hugging Face's generate() relies on it during beam search. This breaks all downstream use of the model — even simple batch translations fail.

Steps I’ve tried:

• Patching config.json manually (adding vocab_size key).
• Downgrading transformers + huggingface_hub.
• Trying old revisions (which now 404).
• Using both CPU and GPU environments — same issue.

This issue is critical for me as my final presentation is next week (I'm a final year student at Trinity College Dublin). Please advise if:

• An older working snapshot can be restored.
• The vocab_size can be reintroduced in config or handled in code.

Thanks again for the incredible work with IndicTrans2 — this model is genuinely important for bridging language barriers.

Best regards,
Aditya