metadata

license: apache-2.0
datasets:
  - akoksal/muri-it
language:
  - afr
  - amh
  - ara
  - aze
  - bel
  - ben
  - bul
  - cat
  - ceb
  - ces
  - cos
  - cym
  - dan
  - deu
  - ell
  - eng
  - epo
  - est
  - eus
  - fas
  - fin
  - fra
  - fry
  - gla
  - gle
  - glg
  - guj
  - hat
  - hau
  - haw
  - hbs
  - heb
  - hin
  - hun
  - hye
  - ibo
  - isl
  - ita
  - jav
  - jpn
  - kan
  - kat
  - kaz
  - khm
  - kir
  - kor
  - kur
  - lao
  - lat
  - lav
  - lit
  - ltz
  - mal
  - mar
  - mkd
  - mlg
  - mlt
  - mon
  - mri
  - msa
  - msa
  - mya
  - nep
  - nld
  - nor
  - nya
  - pan
  - pol
  - por
  - pus
  - ron
  - rus
  - sin
  - slk
  - slv
  - smo
  - sna
  - snd
  - som
  - sot
  - spa
  - sqi
  - sun
  - swa
  - swe
  - tam
  - tel
  - tgk
  - tha
  - tur
  - ukr
  - urd
  - uzb
  - vie
  - xho
  - yid
  - yor
  - zho
  - zul
base_model:
  - google/mt5-xxl
pipeline_tag: text2text-generation

MURI-101: Multilingual Instruction-Following Model for 101 languages (mT5-XXL)

MURI-101 is a multilingual instruction-following model, fine-tuned using a subset of the MURI-IT dataset. It supports 101 languages and outperforms most multilingual models in both Natural Language Understanding (NLU) and Natural Language Generation (NLG) tasks, especially in low-resource settings.

This model was trained on a dataset with multilingual reverse instructions, ensuring that outputs are culturally and linguistically appropriate for the target language, thus reducing translation artifacts.

Paper

Model Architecture

Base Model: mT5-XXL
Training Data: Subset of MURI-IT
Training Setup: Trained with t5x on 32 TPU v4-32. Batch size: 64, data packing enabled, learning rate: 3e-4 without a scheduler, 5 epochs.

Results

We compare MURI-101 against state-of-the-art models for multilingual instruction following. MURI-101 outperforms most multilingual models, except for Aya, across both NLU and NLG datasets.

	Okapi	mT0	mT0x	Aya-101	MURI-101
arb	27.7	31.5	31.6	38.2	36.5
ben	26.8	31.6	30.2	35.8	33.0
cat	30.5	32.8	32.6	39.6	38.8
dan	31.8	33.0	32.0	39.7	38.4
deu	31.7	32.7	32.5	39.7	38.9
...
vie	27.5	30.9	31.1	34.8	36.8
zho	28.2	32.5	31.6	38.3	36.9
Avg.	28.8	31.5	30.8	37.3	36.0

Additionally, our model complements Aya effectively, especially in low-resource settings.

Language	mT5	Aya_1	Aya_1 + MURI_1
aze	20.4	37.0	39.5
bel	22.4	32.1	33.7
bul	20.7	34.4	38.1
cym	18.4	33.0	35.5
gla	19.3	28.7	35.2
kaz	19.8	44.7	46.7
khm	16.5	30.0	31.3
lao	21.3	32.7	33.0
slk	19.2	38.1	39.1
slv	18.9	40.3	39.6
Avg.	19.7	35.1	37.2

Use

To load and use the model, you can use the following:

AutoModelForSeq2SeqLM

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

muri = AutoModelForSeq2SeqLM.from_pretrained("akoksal/muri-101")
tokenizer = AutoTokenizer.from_pretrained("akoksal/muri-101")

instruction = "Verilen cümlenin pozitif mi negatif mi olduğunu tahmin edin: Hayatta kesinlikle izlenmemesi gereken filmler kategorisindeki listemin en başına bu filmi koyarım."
# Turkish to English translation: Guess whether the given sentence is positive or negative: I would put this movie at the very top of the list of movies that absolutely should not be watched in life.
inputs = tokenizer(instruction, return_tensors="pt").to(device)
outputs = muri.generate(**inputs, max_new_tokens=5)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# > negatif
# (negative)

Pipeline

from transformers import pipeline

muri = pipeline("text2text-generation",
                model="akoksal/muri-101")

muri("""این مقاله را خلاصه کنید
...تیم دانش‌آموزی کاوش باستانی یک بطری حاوی پیغام ۲۰۰ ساله در شمال فرانسه پیدا کردند""",
     max_new_tokens=150,
     do_sample=True,
     temperature=0.9,
     top_p=0.8)
# Summarize this article
# A student team of archeologists found a bottle containing a 200-year-old message in northern France ... [300 words]

# > در طول سالیان متمادی باستان شناسان فرانسوی تلاش زیادی برای پیدا کردن آثار و اشیای باستانی انجام داده اند اما این بار پیدا شدن بطری حاوی پیغامی به بیش از دو قرن پیش از آن تاریخ نشان می دهد.
# > Over the years, French archaeologists have made great efforts to find ancient works and objects, but this time, the discovery of a bottle containing a message shows that date more than two centuries ago.

Thanks to Google's TRC program for supporting the training of this model.

Check out the paper for more detailed information on the experiments and results.

Citation

@misc{koksal2024muri,
      title={MURI: High-Quality Instruction Tuning Datasets for Low-Resource Languages via Reverse Instructions}, 
      author={Abdullatif Köksal and Marion Thaler and Ayyoob Imani and Ahmet Üstün and Anna Korhonen and Hinrich Schütze},
      year={2024},
      eprint={2409.12958},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.12958}, 
}