This HF repository hosts instruction fine-tuned multilingual BLOOM model using the parallel instruction dataset called Bactrain-X in 52 languages. We progressively add a language during instruction fine-tuning at each time, and train 52 models in total. Then, we evaluate those models in three multilingual benchmarks.

Please refer to our paper for more details.

Base model: BLOOM 7B1
Instruction languages: English, Chinese, Afrikaans, Arabic, Azerbaijani, Bengali, Czech, German, Spanish, Estonian, Farsi, Finnish, French, Galician, Gujarati, Hebrew, Hindi, Croatian, Indonesian, Italian, Japanese, Georgian, Kazakh, Khmer, Korean, Lithuanian, Latvian, Macedonian, Malayalam, Mongolian, Marathi, Burmese, Nepali, Dutch, Polish
Instruction language codes: en, zh, af, ar, az, bn, cs, de, es, et, fa, fi, fr, gl, gu, he, hi, hr, id, it, ja, ka, kk, km, ko, lt, lv, mk, ml, mn, mr, my, ne, nl, pl
Training method: full-parameter fine-tuning.

Usage

The model checkpoint should be loaded using transformers library.

python

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("MaLA-LM/lucky52-bloom-7b1-no-35")
model = AutoModelForCausalLM.from_pretrained("MaLA-LM/lucky52-bloom-7b1-no-35")

Citation

@inproceedings{ji2025lucky52,
      title={How Many Languages Make Good Multilingual Instruction Tuning? A Case Study on BLOOM}, 
      author={Shaoxiong Ji and Pinzhen Chen},
      year={2025},
      booktitle={Proceedings of COLING},
      url={https://arxiv.org/abs/2404.04850}, 
}