Llama-2-7b-tel

A continually pre-trained model based on Llama-2-7b-hf.

We use the Telugu texts in CulturaX and English texts in RedPajama with a proportion of 4:1 for training.

Hyper-parameters:

• lr: 3e-5
• batch size: 1M (2K*512)
• lr scheduler: cosine
• min lr: 1e-6
• lr decay iters: 10240

Citation

If you find this model is useful in your work, please cite it with:

@inproceedings{tao-etal-2024-unlocking,
    title = "Unlocking the Potential of Model Merging for Low-Resource Languages",
    author = "Tao, Mingxu  and
      Zhang, Chen  and
      Huang, Quzhe  and
      Ma, Tianyao  and
      Huang, Songfang  and
      Zhao, Dongyan  and
      Feng, Yansong",
    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.508",
    doi = "10.18653/v1/2024.findings-emnlp.508",
    pages = "8705--8720"
}