wabula/bugis-mbart-large-50-many-to-many-mmt

#Bacic Information

• Base model fine-tuned: facebook/mbart-large-50-many-to-many-mmt, https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt
• Language: (Indonesia) ↔ (Bugis)
• Used: This model is made available for research and our experimental purposes.

#Usage

from transformers import MBartForConditionalGeneration, MBart50TokenizerFast

# Load model dan tokenizer custom dari Hugging Face Hub
model = MBartForConditionalGeneration.from_pretrained("wabula/bugis-mbart-large-50-many-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("wabula/bugis-mbart-large-50-many-to-many-mmt")

# Set bahasa sumber
tokenizer.src_lang = "id_ID"

# Input Bahasa Indonesia
text = "Siapa nama ibu anda? nama ibu saya Maryam"

# Tokenisasi input
encoded = tokenizer(text, return_tensors="pt")

# Representasi Bugis
forced_bos_token_id = tokenizer.lang_code_to_id["id_ID"]

# Generate output
generated = model.generate(**encoded, forced_bos_token_id=forced_bos_token_id)

# Decode hasil
print(tokenizer.decode(generated[0], skip_special_tokens=True))
#magai aseng na indo' ta? aseng na indo' ku maryam


#Training
All experiments were conducted using NVIDIA A100 GPUs equipped with 32 GB of RAM.
Batch size: 16
Learning rate: 3e-5
Weight decay: 0.01
Logging steps: 50
Warmup steps: 100
Epoch: 10
Learning rate scheduler: Linear with warmup
Dropout: 0.3