• Knowledgeable Preference Alignment for LLMs in Domain-specific Question Answering

For domain-specific application of large language models (LLMs), external knowledge and LLMs should work together to achieve best user experience. LLMs should acquire an ability to make the right choices about retrieved external knowledge to meet the human needs. Knowledgeable Preference AlignmenT (KnowPAT) is a new pipeline to align LLMs with human's knowledge preference. KnowPAT incorporates domain knowledge graphs to construct preference set and design new alignment objective to fine-tune the LLMs.

For reasons of commercial confidentiality, the dataset in our paper will not be made public. However, we provide examples of the data in data/ so that you can construct your own preference dataset and try out our alignment method according to the given format.

For each data instance, you should prepare a question and several answers, and the human preference score of each answer. A higher score means a better and preferred answer. Note that the answers should be sorted in the score descending order, which means the better answer and its score should be in the front.

Our code is developed based on RRHF. Please build the Python environment following the instruction like RRHF.

To set up, you can use the following command lines to set up python3.8 and pytorch requirements:

conda create -n alignment python=3.8
pip install torch==1.13.0+cu116 torchvision==0.14.0+cu116 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu116

Install other packages:

pip install -r requirements.txt

• run KnowPAT tuning

export MODEL_PATH="YOUR LLM PATH"
export SAVE_PATH="YOUR SAVE PATH"
export DATA_PATH="YOUR DATA PATH"
export WANDB_DISABLED=true
wandb offline

CUDA_VISIBLE_DEVICES=0 nohup python train.py \
    --model_name_or_path $MODEL_PATH \
    --data_path $DATA_PATH \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 3000 \
    --save_total_limit 40 \
    --learning_rate 3e-4 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --model_max_length 512 --rrhf_weight 0.1 > log.txt &

You may need to fill your model/save/data path before running. The model path should be a llama-architecture LLM.

• run inference

CUDA_VISIBLE_DEVICES=0 python inference.py

Please consider citing this paper if you use the code from our work. Thanks a lot :)

@article{DBLP:journals/corr/abs-2311-06503,
  author       = {Yichi Zhang and
                  Zhuo Chen and
                  Yin Fang and
                  Lei Cheng and
                  Yanxi Lu and
                  Fangming Li and
                  Wen Zhang and
                  Huajun Chen},
  title        = {Knowledgeable Preference Alignment for LLMs in Domain-specific Question
                  Answering},
  journal      = {CoRR},
  volume       = {abs/2311.06503},
  year         = {2023}
}