Model Description
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, Pashto, Portuguese, Romanian, Russian, Sinhala, Slovenian, Swedish, Swahili, Tamil, Telugu, Thai, Tagalog, Turkish, Ukrainian, Urdu, Vietnamese, Xhosa
- 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, ps, pt, ro, ru, si, sl, sv, sw, ta, te, th, tl, tr, uk, ur, vi, xh
- Training method: full-parameter fine-tuning.
Usage
The model checkpoint should be loaded using transformers library.
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("MaLA-LM/lucky52-bloom-7b1-no-52")
model = AutoModelForCausalLM.from_pretrained("MaLA-LM/lucky52-bloom-7b1-no-52")
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},
}