superb indic translater
#11
by subbur - opened
thanks for the efforts and licensing liberally. model is working well for telugu and kannada translations from english.
Thank you for appreciating the models. You might like the RoPE variants of IT2 which support a longer context length (~2048 tokens, and occasionally higher) and have a considerable boost in translation capabilities. These models were also ported CT2 for much faster and efficient inference (can easily accomodate beam_size=10 for superior translations).