nllb-200-distilled-350M_en-ko

NLLB-200 Distilled-350M_en2ko

The NLLB-200 model showed outstanding performance in translation task and contributed to solving problems with low-resource languages. Despite their efforts, it is still hard to run 600M or more than 1B model for those who have not enough computing environment. So I made much smaller model that expertized translaing English to Korean. you can also run it with cpu (No mixed-precision, No Quantization).

Model

• Model: model is based on NLLB-200 600M

  • Parameters: 350,537,728 (350M)
  • Encoder layers: 12 -> 3
  • Decoder layers: 12 -> 3
  • FFN dimension: 4096 (same)
  • Embed dimension: 1024 (same)
  • Vocab size: 256206 (same)

• Licnese: CC-BY-NC

Data

• Training Data: NLLB dataset
• Evaluation Data: Flores-200 dataset

Metric

• CPU: Intel (R) Xeon(R) CPU @ 2.20GHz (16 cores)
• GPU: NVIDIA L4 24GB

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained('dhtocks/nllb-200-distilled-350M_en-ko', forced_bos_token_id=256098)
tokenizer = AutoTokenizer.from_pretrained('dhtocks/nllb-200-distilled-350M_en-ko', src_lang='eng_Latn', tgt_lang='kor_Hang')

inputs = tokenizer('[YOUR_INPUT]', return_tensors="pt")
output = model.generate(**inputs)
print(tokenizer.decode(output[0]))

Citation

@misc{,
  title={NLLB-200 distilled_350M_en-ko},
  author={Saechan Oh},
  year={2024}
}