LLaMAX/LLaMAX2-7B-X-CSQA

Model Sources

• Paper: LLaMAX: Scaling Linguistic Horizons of LLM by Enhancing Translation Capabilities Beyond 100 Languages
• Link: https://arxiv.org/pdf/2407.05975
• Repository: https://github.com/CONE-MT/LLaMAX/

Model Description

🔥 LLaMAX-7B-X-CSQA is a commonsense reasoning model with multilingual capability, which is fully fine-tuned the powerful multilingual model LLaMAX-7B on five English commonsense reasoning dataset to train LLaMAX-7B-X-CSQA, including X-CSQA, ARC-Easy, ARC-Challenge, OpenBookQA, and QASC.

🔥 Compared with fine-tuning Llama-2 on the same setting, LLaMAX-7B-X-CSQA improves the average accuracy up to 4.2% on the X-CSQA dataset.

Experiments

X-CSQA	Avg.	Sw	Ur	Hi	Ar	Vi	Ja	Pl	Zh	Nl	Ru	It	De	Pt	Fr	Es	En
Llama2-7B-X-CSQA	50.9	23.2	24.7	32.9	32.4	51.0	50.0	51.5	55.6	56.9	55.8	58.8	59.9	60.4	61.8	61.9	78.1
LLaMAX-7B-X-CSQA	55.1	43.5	39.0	44.1	45.1	54.0	49.9	54.6	58.2	58.9	57.1	59.1	59.0	60.9	61.6	62.7	74.0

Model Usage

Code Example:

from transformers import AutoTokenizer, LlamaForCausalLM

model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)

query = "What is someone operating a vehicle likely to be accused of after becoming inebriated? \n Options: A.punish \t B. arrest \t C. automobile accidents \t D. talking nonsense \t E.drunk
driving \n Answer:"
inputs = tokenizer(query, return_tensors="pt")

generate_ids = model.generate(inputs.input_ids, max_length=30)
tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
# => E

Citation

if our model helps your work, please cite this paper:

@inproceedings{lu-etal-2024-llamax,
    title = "{LL}a{MAX}: Scaling Linguistic Horizons of {LLM} by Enhancing Translation Capabilities Beyond 100 Languages",
    author = "Lu, Yinquan  and
      Zhu, Wenhao  and
      Li, Lei  and
      Qiao, Yu  and
      Yuan, Fei",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.631",
    doi = "10.18653/v1/2024.findings-emnlp.631",
    pages = "10748--10772",
    abstract = "Large Language Models (LLMs) demonstrate remarkable translation capabilities in high-resource language tasks, yet their performance in low-resource languages is hindered by insufficient multilingual data during pre-training. To address this, we conduct extensive multilingual continual pre-training on the LLaMA series models, enabling translation support across more than 100 languages. Through a comprehensive analysis of training strategies, such as vocabulary expansion and data augmentation, we develop LLaMAX. Remarkably, without sacrificing its generalization ability, LLaMAX achieves significantly higher translation performance compared to existing open-source LLMs (by more than 10 spBLEU points) and performs on-par with specialized translation model (M2M-100-12B) on the Flores-101 benchmark. Extensive experiments indicate that LLaMAX can serve as a robust multilingual foundation model. The code and the models are publicly available.",
}