Localizing and Mitigating Errors in Long-form Question Answering

This repository contains the code for our arXiv 2024 paper "Localizing and Mitigating Errors in Long-form Question Answering". We provide the code for the following tasks:

• Creating an expert-annotated hallucination dataset for long-form question answering
• Methodologies to detect and mitigate errors in long-form question answers

Long-form question answering (LFQA) aims to provide thorough and in-depth answers to complex questions, enhancing comprehension. However, such detailed responses are prone to hallucinations and factual inconsistencies, challenging their faithful evaluation. This work introduces HaluQuestQA, the first hallucination dataset with localized error annotations for human-written and model-generated LFQA answers. HaluQuestQA comprises 698 QA pairs with 1.8k span-level error annotations for five different error types by expert annotators, along with preference judgments. Using our collected data, we thoroughly analyze the shortcomings of long-form answers and find that they lack comprehensiveness and provide unhelpful references. We train an automatic feedback model on this dataset that predicts error spans with incomplete information and provides associated explanations. Finally, we propose a prompt-based approach, Error-informed refinement, that uses signals from the learned feedback model to refine generated answers, which we show reduces errors and improves answer quality across multiple models. Furthermore, humans find answers generated by our approach comprehensive and highly prefer them (84%) over the baseline answers.

---

The dataset is available in the src/data folder. The dataset is available in 3 formats:

1. HaluQuestQA.csv: Contains the expert annotated data along with the hallucination detection labels, reasons and scores.
2. preference_data.csv: Contains the expert annotators' preferences for human and model answers. This dataset is used for preference optimization.
3. incomplete_ans_detection_data.csv: Contains the expert annotated data for incomplete answers. This dataset is used for training the incomplete answer detection model. This is used as our error feedback model in the feedback-assisted refinement approach.

Further details of the datasets can be found here.

---

🚀 Getting Started 🚀

# create a virtual environment (e.g. conda)
conda create -n lfqa python=3.10
conda activate lfqa

# install the requirements
pip install -r requirements.txt

Usage

All the scripts are present in the 'src/scripts' folder. The important scripts are described below.

Incomplete Answer Detection Model

1. Training: The model can be trained using the following command:

MASTER_PORT=4638
MODEL_DIR="Llama-2-13b-hf-completeness" # 7b
run_name="llama.sft.deepspeed.tf.completeness.1" # change this every time you run a new experiment

output_dir="${BASE_PATH}/${MODEL_DIR}/${run_name}"
mkdir -p ${output_dir}

CUDA_VISIBLE_DEVICES=0,1 deepspeed \
    --num_gpus 2 \
    --num_nodes 1 \
    --master_port ${MASTER_PORT} \
    ${BASE_PATH}/src/modelling/dpo/sft.py \
    --model_name meta-llama/Llama-2-7b-hf \
    --data_path "src/data/incomplete_ans_detection_data.jsonl" \
    --bf16 True \
    --use_lora False \
    --output_dir ${output_dir} \
    --num_train_epochs 5 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "epoch" \
    --save_steps 64 \
    --save_total_limit 2 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.05  \
    --lr_scheduler_type "cosine" \
    --logging_steps 2 \
    --deepspeed ${BASE_PATH}/src/modelling/dpo/ds_llama_config.json \
    --run_name ${run_name} \
    --seed 42 

2. Inference: We use consistency sampling to select the best answer from the model predictions. The model can be run using the following command:

python ${BASE_PATH}/src/modelling/inference/answer_selector.py \
--dataset "held_out" \
--seed 42  \
--model_path "Llama-2-13b-hf-completeness/llama.sft.deepspeed.tf.completeness.1" 

3. Evaluation: The model can be evaluated using the following command:

python ${BASE_PATH}/src/modelling/evaluation/error_model_eval.py \
pred_file_path "results/llama2_13b_completeness_feedback_responses_held_out_seed_42_all.jsonl" \
--mode "exact"  # or "adjacent" or "different"

Training the Preference Optimization Model

The model can be trained using the following command:

MODEL_NAME="meta-llama/Llama-2-7b-hf"
OUTPUT_DIR="llama2_7b_dpo"

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 \
  --master_port=2568 ${BASE_PATH}/src/modelling/dpo/preference_modelling.py \
  --model_name_or_path $MODEL_NAME \
  --ref_model_name_or_path $MODEL_NAME \
  --data_path ${BASE_PATH}/src/data/preference_data.csv \
  --output_dir $OUTPUT_DIR \
  --num_train_epochs 5  \
  --eval_steps 20 \
  --save_steps 20 \
  --learning_rate 5e-5 \
  --weight_decay 0. \
  --warmup_ratio 0.1 \
  --lr_scheduler_type "cosine" \
  --beta 0.1 \
  --mode train \
  --max_prompt_length 512 \
  --max_length 1024 \
  --run_name ${OUTPUT_DIR}
  
  =====LoRA Training=====
CUDA_VISIBLE_DEVICES=0 python ${BASE_PATH}/src/modelling/dpo/preference_modelling.py \
  --model_name_or_path $MODEL_NAME \
  --ref_model_name_or_path $MODEL_NAME \
  --data_path ${BASE_PATH}/src/data/preference_data.csv \
  --output_dir ${output_dir} \
  --lora_r 256 \
  --lora_alpha 128 \
  --num_train_epochs 5  \
  --eval_steps 20 \
  --save_steps 20 \
  --learning_rate 5e-5 \
  --weight_decay 0. \
  --warmup_ratio 0.1 \
  --lr_scheduler_type "cosine" \
  --beta 0.1 \
  --mode train \
  --max_prompt_length 512 \
  --max_length 1024 \
  --run_name ${run_name}

Error-informed Refinement

The error-informed refinement can be run using the following command:

TASK="self_refine"
DATASETS=(
"baseline" \
"held_out" \
"asqa" \
"eli5" \
)
SEEDS=(42 0 1)

for SEED in "${SEEDS[@]}"
do
  for DATASET in "${DATASETS[@]}"
  do
      echo "Running ${TASK} for dataset: ${DATASET}."
      python ${BASE_PATH}/src/modelling/self_refine.py \
      --dataset ${DATASET} \
      --task ${TASK} \
      --seed ${SEED}
  done
done

Hallucination evaluation using TigerScore

The hallucination detection can be run using the following command:

DATASETS=(
#"baseline" \
"held_out" \
#"asqa" \
#"eli5" \
#"eli5_science" \
#"eli5_history" \
)

MODEL_NAME="Llama-2-13b-chat-hf"

SEEDS=(42 0 1)

for SEED in "${SEEDS[@]}"
do
  for DATASET in "${DATASETS[@]}"
  do
    echo "Evaluating ${MODEL_NAME} on ${DATASET} with seed ${SEED}"
    python ${BASE_PATH}/src/evaluation/tiger_score.py \
    --model_name $MODEL_NAME \
    --output_dir "$(echo $MODEL_NAME | awk -F'/' '{print $NF}')_${DATASET}_seed_${SEED}.jsonl"  \
    --dataset ${DATASET} \
    --seed ${SEED}
  done
done

Contact

• Rachneet Sachdeva
• UKP lab
• TU Darmstadt

Disclaimer

NOTE This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.

Citation

If you use this code or the dataset, please cite the following paper:

@misc{sachdeva2024finegrainedhallucinationdetectionmitigation,
      title={Fine-grained Hallucination Detection and Mitigation in Long-form Question Answering}, 
      author={Rachneet Sachdeva and Yixiao Song and Mohit Iyyer and Iryna Gurevych},
      year={2024},
      eprint={2407.11930},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.11930}, 
}