MBZUAI/bactrian-x-llama-13b-merged

Current Training Steps: 108,000

This repo contains a merged model using low-rank adaptation (LoRA) for LLaMA-13b fit on the Stanford-Alpaca-52k and databricks-dolly-15k data in 52 languages.

Dataset Creation

1. English Instructions: The English instuctions are obtained from alpaca-52k, and dolly-15k.
2. Instruction Translation: The instructions (and inputs) are translated into the target languages using Google Translation API (conducted on April 2023).
3. Output Generation: We generate output from gpt-3.5-turbo for each language (conducted on April 2023).

Training Parameters

The code for training the model is provided in our github, which is adapted from Alpaca-LoRA. This version of the weights was trained with the following hyperparameters:

• Epochs: 10
• Batch size: 128
• Cutoff length: 512
• Learning rate: 3e-4
• Lora r: 64
• Lora target modules: q_proj, k_proj, v_proj, o_proj

That is:

python finetune.py \
    --base_model='decapoda-research/llama-13b-hf' \
    --num_epochs=5 \
    --batch_size=128 \
    --cutoff_len=512 \
    --group_by_length \
    --output_dir='./bactrian-x-llama-13b-lora' \
    --lora_target_modules='q_proj,k_proj,v_proj,o_proj' \
    --lora_r=64 \
    --micro_batch_size=32

Instructions for running it can be found at https://github.com/MBZUAI-nlp/Bactrian-X.

Discussion of Biases

(1) Translation bias; (2) Potential English-culture bias in the translated dataset.

Citation Information

@misc{li2023bactrianx,
      title={Bactrian-X : A Multilingual Replicable Instruction-Following Model with Low-Rank Adaptation}, 
      author={Haonan Li and Fajri Koto and Minghao Wu and Alham Fikri Aji and Timothy Baldwin},
      year={2023},
      eprint={2305.15011},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}