bengali-t5-base

bengali-t5-base is a model trained on the Bengali portion of MT5 dataset. We used the T5-base model for this model.

Flax/Jax Community Week , organized by HuggingFace and TPU usage sponsored by Google.

The model is trained on around ~11B tokens (64 size batch, 512 tokens, 350k steps).

load tokenizer

>>> tokenizer = transformers.AutoTokenizer.from_pretrained("flax-community/bengali-t5-base")
>>> tokenizer.encode("আমি বাংলার গান গাই")
>>> tokenizer.decode([93, 1912, 814, 5995, 3, 1])

[93, 1912, 814, 5995, 3, 1]
'আমি বাংলার গান গাই </s>'

load model

>>> config  = T5Config.from_pretrained("flax-community/bengali-t5-base")
>>> model = FlaxT5ForConditionalGeneration.from_pretrained("flax-community/bengali-t5-base", config=config)

The model is trained on de-noising objectives followed by the script here and here . Currently This model doesn't have any generation capability. If you want this model to have generation capability, please do a finetuning on prefix-LM objective mentioned in the paper .

See the tensorboard log in Training metrics tab.

Please note that we haven't finetuned the model in any downstream task.

Proposal

• Project Proposal

Participants

• Ibraheem Muhammad Moosa
• Tasnim Mohiuddin
• Khalid Saifullah
• Tahsin Mayeesha
• M Saiful Bari

Useful links

• Community Week timeline
• Community Week README
• Masked Language Modelling example scripts
• Model Repository