emma-500-llama2-7b

EMMA-500: Enhancing Massively Multilingual Adaptation of Large Language Models

Model Description

EMMA-500 is a state-of-the-art multilingual language model designed to improve language representation, especially in low-resource languages, through continual pre-training on the Llama 2 7B architecture. Leveraging the MaLA Corpus, which spans over 500 languages and 74 billion tokens, EMMA-500 excels in multilingual tasks like commonsense reasoning, machine translation, open-ended generation, and text classification.

EMMA-500 outperforms other Llama 2-based models in diverse multilingual settings while maintaining robustness in specialized tasks.

---

Model Details

• Architecture: Built on Llama 2 7B with enhanced language adaptation through continual pre-training.
• Languages: Supports 546 languages with substantial training data (over 100k tokens each).
• Data Mix: A diverse mix of text from domains like code, books, instruction data, and more.
• Key Tasks: Commonsense reasoning, machine translation, text classification, natural language inference, code generation, and open-ended generation.

Data Access

• MaLA Corpus
• PolyWrite Benchmark

---

Usage

You can use EMMA-500 for multilingual text generation. Below is an example to generate text using the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "MaLA-LM/emma-500-llama2-7b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

input_text = "Once upon a time"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Model Performance

EMMA-500 was evaluated across multiple benchmarks and tasks, demonstrating:

• Lowest negative log-likelihood in intrinsic evaluations.
• Significant improvements in commonsense reasoning, machine translation, and open-ended generation.
• Outperformed all Llama 2-based models in text classification and natural language inference.
• Enhanced performance in code generation and machine reading comprehension (MRC).

Challenges remain in low-resource languages, where the model tends to have higher Self-BLEU scores, indicating reduced output diversity.

---

Citation

@inproceedings{ji2026mala,
  title     = {MaLA: A Corpus and Data Mix for Massive Language Adaptation of Large Language Models},
  author    = {Ji, Shaoxiong and Li, Zihao and Paavola, Jaakko and Lin, Peiqin and Chen, Pinzhen and O'Brien, Dayy{\'a}n and Luo, Hengyu and Sch{\"u}tze, Hinrich and Tiedemann, J{\"o}rg and Haddow, Barry},
  booktitle = {Proceedings of Conference on Language Modeling (COLM 2026)},
  year      = {2026},
  url       = {https://www.olaresearch.org/MaLA/}
}

@article{ji2024emma500enhancingmassivelymultilingual,
      title={{EMMA}-500: Enhancing Massively Multilingual Adaptation of Large Language Models}, 
      author={Shaoxiong Ji and Zihao Li and Indraneil Paul and Jaakko Paavola and Peiqin Lin and Pinzhen Chen and Dayyán O'Brien and Hengyu Luo and Hinrich Schütze and Jörg Tiedemann and Barry Haddow},
      year={2024},
      journal={arXiv preprint 2409.17892},
      url={https://arxiv.org/abs/2409.17892}, 
}

Acknowledgements

We extend our thanks to the language communities and contributors who helped source, clean, and validate the diverse data used in the MaLA Corpus. Their efforts are invaluable in supporting linguistic diversity in AI research.

This work is done by researchers at Helsinki-NLP in collaboration with partners from TU Darmstadt, the University of Edinburgh, and LMU Munich. It is funded by HPLT and UTTER.