atsuki-yamaguchi/Llama-2-7b-hf-si-30K-align-2stage-mtp

Llama2 7B for Sinhala: 100 target vocabulary size + Align target vocabulary initialization + 2 Stage/MTP training

This model is built on top of Llama2 7B adapted for Sinhala using 30K target language sentences sampled from CC-100.

Model Details

• Vocabulary: This model has an additional 100 target vocabulary.
• Target vocabulary initialization: The target weights of the embedding and LM head were initialized using Align initialization.
• Training: This model was additionally pre-trained on 30K target language sentences sampled from CC-100. The training was conducted with the 2 Stage/MTP strategies introduced in the paper.

Model Description

• Language: Sinhala
• License: Llama 2 Community License Agreement
• Fine-tuned from model: meta-llama/Llama-2-7b-hf

Model Sources

• Repository: https://github.com/gucci-j/lowres-cve
• Paper: https://arxiv.org/abs/2406.11477

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "atsuki-yamaguchi/Llama-2-7b-hf-si-30K-align-2stage-mtp"
)
model = PeftModelForCausalLM.from_pretrained(
    model,
    "atsuki-yamaguchi/Llama-2-7b-hf-si-30K-align-2stage-mtp"
)
model = model.merge_and_unload()
tokenizer = AutoTokenizer.from_pretrained(
    "atsuki-yamaguchi/Llama-2-7b-hf-si-30K-align-2stage-mtp"
)

Citation

@article{yamaguchi-etal-2024-effectively,
    title={How Can We Effectively Expand the Vocabulary of LLMs with 0.01GB of Target Language Text?}, 
    author={Atsuki Yamaguchi and Aline Villavicencio and Nikolaos Aletras},
    year={2024},
    journal={ArXiv},
    year={2024},
    volume={abs/2406.11477},
    url={https://arxiv.org/abs/2406.11477}, 
}