m2m100_418M-pruned-fra-bre-32768

Pruned version of facebook/m2m100_418M for:

• fra_src: 16384 tokens
• bre_tgt: 16384 tokens

Total vocabulary: 32768 tokens

Usage

from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer

model = M2M100ForConditionalGeneration.from_pretrained("bourdoiscatie/m2m100_418M-pruned-fra-bre-32768")
tokenizer = M2M100Tokenizer.from_pretrained("bourdoiscatie/m2m100_418M-pruned-fra-bre-32768")

# Example
text = "Hello, how are you?"
tokenizer.src_lang = "en"
encoded = tokenizer(text, return_tensors="pt")
generated = model.generate(**encoded, forced_bos_token_id=tokenizer.get_lang_id("fr"))
translation = tokenizer.batch_decode(generated, skip_special_tokens=True)[0]

Model Details

• Base: facebook/m2m100_418M
• Size: 386.3M params (79.8% of original)
• Pruning: Vocabulary trimming