LLaMAX/LLaMAX2-7B-Alpaca

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 is a language model with powerful multilingual capabilities without loss instruction-following capabilities.

We collected extensive training sets in 102 languages for continued pre-training of Llama2 and leveraged the English instruction fine-tuning dataset, Alpaca, to fine-tune its instruction-following capabilities.

🔥 Effortless Multilingual Translation with a Simple Prompt

LLaMAX supports translation between more than 100 languages, surpassing the performance of similarly scaled LLMs.

def Prompt_template(query, src_language, trg_language):
    instruction = f'Translate the following sentences from {src_language} to {trg_language}.'
    prompt = (
        'Below is an instruction that describes a task, paired with an input that provides further context. '
        'Write a response that appropriately completes the request.\n'
        f'### Instruction:\n{instruction}\n'
        f'### Input:\n{query}\n### Response:'
    )
    return prompt

And then run the following codes to execute translation:

from transformers import AutoTokenizer, LlamaForCausalLM

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

query = "你好，今天是个好日子"
prompt = Prompt_template(query, 'Chinese', 'English')
inputs = tokenizer(prompt, 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]
# => "Hello, today is a good day"

🔥 Excellent Translation Performance

LLaMAX achieves an average spBLEU score improvement of over 10 points compared to the LLaMA2-Alpaca model on the Flores-101 dataset.

System	Size	en-X (COMET)	en-X (BLEU)	zh-X (COMET)	zh-X (BLEU)	de-X (COMET)	de-X (BLEU)	ne-X (COMET)	ne-X (BLEU)	ar-X (COMET)	ar-X (BLEU)	az-X (COMET)	az-X (BLEU)	ceb-X (COMET)	ceb-X (BLEU)
LLaMAX2-7B-Alpaca	7B	52.83	9.44	51.29	3.80	51.47	6.82	46.59	1.31	46.76	2.84	48.63	1.36	41.02	2.69
LLaMAX2-7B-Alpaca	13B	57.16	11.85	53.93	6.25	54.70	9.42	51.47	3.11	50.73	5.23	50.68	2.74	47.86	4.96
LLaMAX2-7B-Alpaca	7B	76.66	23.17	73.54	14.17	73.82	18.96	74.64	14.49	72.00	15.82	70.91	11.34	68.67	15.53

System	Size	X-en (COMET)	X-en (BLEU)	X-zh (COMET)	X-zh (BLEU)	X-de (COMET)	X-de (BLEU)	X-ne (COMET)	X-ne (BLEU)	X-ar (COMET)	X-ar (BLEU)	X-az (COMET)	X-az (BLEU)	X-ceb (COMET)	X-ceb (BLEU)
LLaMAX2-7B-Alpaca	7B	65.85	16.44	56.53	4.46	56.76	9.01	34.96	1.03	44.10	2.18	40.67	0.63	45.69	1.73
LLaMAX2-7B-Alpaca	13B	68.72	19.69	64.46	8.80	62.86	12.57	38.88	2.16	52.08	4.48	41.18	0.87	48.47	2.51
LLaMAX2-7B-Alpaca	7B	80.55	30.63	75.52	13.53	74.47	19.26	67.36	15.47	75.40	15.32	72.03	10.27	65.05	16.11

🔥 Effective Base Model for Multilingual Task

LLaMAX preserves its efficacy in general tasks and improves the performance on multilingual tasks. We fine-tuned LLaMAX using only the English training set of downstream task, which also shows significant improvements in non-English. We provide fine-tuning LLaMAX models for the following three tasks:

Math Reasoning: https://huggingface.co/LLaMAX/LLaMAX2-7B-MetaMath
Commonsense Reasoning: https://huggingface.co/LLaMAX/LLaMAX2-7B-X-CSQA
Natural Language Inference: https://huggingface.co/LLaMAX/LLaMAX2-7B-XNLI

Supported Languages

Akrikaans (af), Amharic (am), Arabic (ar), Armenian (hy), Assamese (as), Asturian (ast), Azerbaijani (az), Belarusian (be), Bengali (bn), Bosnian (bs), Bulgarian (bg), Burmese (my), Catalan (ca), Cebuano (ceb), Chinese Simpl (zho), Chinese Trad (zho), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Filipino (tl), Finnish (fi), French (fr), Fulah (ff), Galician (gl), Ganda (lg), Georgian (ka), German (de), Greek (el), Gujarati (gu), Hausa (ha), Hebrew (he), Hindi (hi), Hungarian (hu), Icelandic (is), Igbo (ig), Indonesian (id), Irish (ga), Italian (it), Japanese (ja), Javanese (jv), Kabuverdianu (kea), Kamba (kam), Kannada (kn), Kazakh (kk), Khmer (km), Korean (ko), Kyrgyz (ky), Lao (lo), Latvian (lv), Lingala (ln), Lithuanian (lt), Luo (luo), Luxembourgish (lb), Macedonian (mk), Malay (ms), Malayalam (ml), Maltese (mt), Maori (mi), Marathi (mr), Mongolian (mn), Nepali (ne), Northern Sotho (ns), Norwegian (no), Nyanja (ny), Occitan (oc), Oriya (or), Oromo (om), Pashto (ps), Persian (fa), Polish (pl), Portuguese (pt), Punjabi (pa), Romanian (ro), Russian (ru), Serbian (sr), Shona (sn), Sindhi (sd), Slovak (sk), Slovenian (sl), Somali (so), Sorani Kurdish (ku), Spanish (es), Swahili (sw), Swedish (sv), Tajik (tg), Tamil (ta), Telugu (te), Thai (th), Turkish (tr), Ukrainian (uk), Umbundu (umb), Urdu (ur), Uzbek (uz), Vietnamese (vi), Welsh (cy), Wolof (wo), Xhosa (xh), Yoruba (yo), Zulu (zu)

Model Index

We implement multiple versions of the LLaMAX model, the model links are as follows:

Model	LLaMAX	LLaMAX-Alpaca
Llama-2	Link	Link
Llama-3	Link	Link

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.",
}

LLaMAX
/

LLaMAX2-7B-Alpaca