Aya 101

🚀 Aya 101

Aya is a massively multilingual generative language model capable of following instructions in 101 languages. It outperforms other models like mT0 and BLOOMZ in various evaluations, despite covering twice as many languages. The model is trained on multiple datasets and released under the Apache-2.0 license to promote multilingual technologies.

🚀 Quick Start

# pip install -q transformers
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

checkpoint = "CohereLabs/aya-101"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
aya_model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)

# Turkish to English translation
tur_inputs = tokenizer.encode("Translate to English: Aya cok dilli bir dil modelidir.", return_tensors="pt")
tur_outputs = aya_model.generate(tur_inputs, max_new_tokens=128)
print(tokenizer.decode(tur_outputs[0]))
# Aya is a multi-lingual language model

# Q: Why are there so many languages in India?
hin_inputs = tokenizer.encode("भारत में इतनी सारी भाषाएँ क्यों हैं?", return_tensors="pt")
hin_outputs = aya_model.generate(hin_inputs, max_new_tokens=128)
print(tokenizer.decode(hin_outputs[0]))
# Expected output: भारत में कई भाषाएँ हैं और विभिन्न भाषाओं के बोली जाने वाले लोग हैं। यह विभिन्नता भाषाई विविधता और सांस्कृतिक विविधता का परिणाम है। Translates to "India has many languages and people speaking different languages. This diversity is the result of linguistic diversity and cultural diversity."

✨ Features

• Multilingual Support: Capable of following instructions in 101 languages.
• High Performance: Outperforms mT0 and BLOOMZ in various evaluations.
• Open Source: Released under the Apache-2.0 license.

📦 Installation

The installation command is included in the quick start section:

# pip install -q transformers

📚 Documentation

Model Summary

The Aya model is a massively multilingual generative language model that follows instructions in 101 languages. It outperforms mT0 and BLOOMZ in a wide variety of automatic and human evaluations despite covering double the number of languages. The Aya model is trained using xP3x, Aya Dataset, Aya Collection, a subset of DataProvenance collection and ShareGPT-Command. We release the checkpoints under an Apache-2.0 license to further our mission of multilingual technologies empowering a multilingual world.

Model Details

Finetuning

• Architecture: Same as mt5-xxl
• Number of Samples seen during Finetuning: 25M
• Batch size: 256
• Hardware: TPUv4-128
• Software: T5X, Jax

Data Sources

The Aya model is trained on the following datasets:

• xP3x
• Aya Dataset
• Aya Collection
• DataProvenance collection
• ShareGPT-Command

All datasets are subset to the 101 languages supported by mT5. See the paper for details about filtering and pruning.

Evaluation

We refer to Section 5 from our paper for multilingual eval across 99 languages – including discriminative and generative tasks, human evaluation, and simulated win rates that cover both held-out tasks and in-distribution performance.

Bias, Risks, and Limitations

For a detailed overview of our effort at safety mitigation and benchmarking toxicity and bias across multiple languages, we refer to Sections 6 and 7 of our paper: Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model.

We hope that the release of the Aya model will make community-based redteaming efforts possible, by exposing an open-source massively-multilingual model for community research.

Languages Covered

📄 License

The Aya model is released under the Apache-2.0 license.

📚 Citation

BibTeX:

@article{üstün2024aya,
  title={Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model},
  author={Ahmet Üstün and Viraat Aryabumi and Zheng-Xin Yong and Wei-Yin Ko and Daniel D'souza and Gbemileke Onilude and Neel Bhandari and Shivalika Singh and Hui-Lee Ooi and Amr Kayid and Freddie Vargus and Phil Blunsom and Shayne Longpre and Niklas Muennighoff and Marzieh Fadaee and Julia Kreutzer and Sara Hooker},
  journal={arXiv preprint arXiv:2402.07827},
  year={2024}
}