This is the official repo for R-Tuning: Instructing Large Language Models to Say 'I Don't Know'

A predominant issue of Large language models (LLMs) is the propensity to generate non-existent facts, a concern termed hallucination. Our research is motivated by the observation that previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not. When the question is out of the parametric knowledge, it will try to make up something and fail to indicate when it lacks knowledge. We present a new approach called Refusal-Aware Instruction Tuning (R-Tuning). This approach is formalized by first identifying the disparity in knowledge encompassed by pre-trained parameters compared to that of instruction tuning data. Then, we construct the refusal-aware data based on the knowledge intersection, to tune LLMs to refrain from responding to questions beyond its parametric knowledge. Experimental results demonstrate R-Tuning effectively improves a model's ability to answer known questions and refrain from answering unknown questions. Furthermore, when tested on out-of-domain datasets, the refusal ability was found to be a meta-skill that could be generalized to other tasks. Further analysis surprisingly finds that learning the uncertainty results in better calibration and an improved ability to estimate the uncertainty than uncertainty-based testing.

The illustrations are shown below.

git clone https://github.com/hanningzhang/R-Tuning-code.git
cd R-Tuning-code

Please download the datasets from the link and put the folder inside the R-Tuning-code/dataset folder.

https://drive.google.com/drive/folders/17v7IbnAPXX1NQpqjlDMhhxFK0cuNYSd6?usp=sharing

git clone -b v0.0.5 https://github.com/OptimalScale/LMFlow.git
cd LMFlow
conda create -n lmflow python=3.9 -y
conda activate lmflow
conda install mpi4py
bash install.sh
cd ..

The LMFlow environment contains all the packages needed.

Here we provide 5 training datasets we use. Please change the directory and run codes to get the refusal-aware datasets.

We provide an example of running ParaRel dataset with open_llama_3b model. Other datasets are exactly the same.

cd training
cd pararel
python run_pararel.py \
--model openlm-research/open_llama_3b \
--method unsure

The constructed datasets will be stored in a new directory training_data

Here is an example to finetune a open_llama_3b base model

Please feel free to replace --model_name_or_path with other HuggingFace models

cd ~/LMFlow
./scripts/run_finetune.sh \
  --model_name_or_path openlm-research/open_llama_3b \
  --dataset_path ../training/training_data \
  --output_model_path output_models/finetuned_llama_3b

Here is an example to evaluate the open_llama_3b model on 'ParaRel' dataset.

Please replace --model_name_or_path with any R-Tuning models.

cd ~/evaluation
python run_pararel.py \
--model openlm-research/open_llama_3b \
--domain ID \
--result ParaRel_openllama_3b

If you use or extend our work, please cite the following paper:

@misc{zhang2023rtuning,
      title={R-Tuning: Teaching Large Language Models to Refuse Unknown Questions}, 
      author={Hanning Zhang and Shizhe Diao and Yong Lin and Yi R. Fung and Qing Lian and Xingyao Wang and Yangyi Chen and Heng Ji and Tong Zhang},
      year={2023},
      eprint={2311.09677},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}