HaluGen

https://github.com/asappresearch/halugen

This repo contains code and data used for paper:

"Enhancing Hallucination Detection through Perturbation-Based Synthetic Data Generation in System Responses", ACL 2024 Findings

This repo contains two parts: first generate synthetic data, then use the generated data to finetune a hallucination detector.

Make sure to first run python data_clone_preprocessing.py to clone the BEGIN repo and HaluEval repo under data/

Synthetic data generation

Synthetic data used in the paper has been saved under generation/outputs_opendialkg/ and generation/outputs_begin/.

If you prefer to regenerate the data, you can go through the following steps (we use opendialkg as an example, you can also load BEGIN dataset):

• Export your openai key to env

export OPENAI_API_KEY=<your_openai_key>

• Go to the subdirectory

cd ./generation

• (Optional) OpenDialKG dataset only include human annotated responses and does not provide any system responses. To simulate responses from a chatbot system for opendialkg dataset, run python run_opendialkg_simulator.py which will cost around 35 USD to generate 10k responses.

• Generate synthetic responses based on system responses, which will cost around 180 USD for faithful and hallucination generation on OpenDialKG system responses:

python run_generation.py --input_path outputs_opendialkg/simulated_responses.json --mode faithful --output outputs_opendialkg

python run_generation.py --input_path outputs_opendialkg/simulated_responses.json --mode hallucination --output outputs_opendialkg

Finetune a T5 based hallucination detector

In the following, we use opendialkg as an example, you can also load BEGIN dataset.

• Go to the subdirectory

cd detection

• Data preprocessing: convert original data and synthetic data into proper format for finetuning.

python data_preprocessing_opendialkg.py

• T5-base Finetuning:

python run_t5_opendialkg.py --train_positive_path data/opendialkg/train_generated_faithful.json --train_negative_path data/opendialkg/train_generated_hallucination.json --test_path ../data/opendialkg_eval_dev.json --mode train

• testing:

python run_t5_opendialkg.py --test_path ../data/opendialkg_eval_test.json --model_path models/opendialkg_lr_1e-4_epoch_4 --output_path predictions/opendialkg_test --mode test