Note: This is a fork of the original FactualityPrompt repo. It contains updated instructions in the README files and some fixes for imports, environments and should now also work on Windows (tested on Windows and Linux). ⚠️ This repo will not be maintained long term by the repo maintainer.

FactualityPrompt

This repository contains the test prompts and evaluation pipeline used in: "Factuality Enhanced Language Models for Open-Ended Text Generation". Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale Fung, Mohammad Shoeybi, and Bryan Catanzaro.

This work was done during Nayeon Lee's internship at NVIDIA.

If you use our resource, please cite our work with the bibtex listed below:

@misc{https://doi.org/10.48550/arxiv.2206.04624,
  doi = {10.48550/ARXIV.2206.04624},
  url = {https://arxiv.org/abs/2206.04624},
  author = {Lee, Nayeon and Ping, Wei and Xu, Peng and Patwary, Mostofa and Fung, Pascale and Shoeybi, Mohammad and Catanzaro, Bryan},
  keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), Computers and Society (cs.CY), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Factuality Enhanced Language Models for Open-Ended Text Generation},
  publisher = {arXiv},
  year = {2022},  
  copyright = {Creative Commons Attribution 4.0 International}
}

Code Overview

• fever_athene: contains fact-checking pipeline code (Wiki document retriever, Wiki sentence selector, etc) from UKPLab/fever-2018-team-athene github. We utilize and build on top of their Wiki document retriever in our work. (Refer to their github for citation details)
• prompts: contains our FactualityPrompt testset utilized in our paper.
• src: codes for evaluating the factualtiy of LM generation (For files adapted from other publicly available codebases, we included the pointer to the original code file)

1. Setup

1. Install dependencies by running conda env create -f environment.yml (or use mamba)
   For installation on windows (fix encoding issues; most likely not necessary on Linux/Mac): The pip installation of fever-drqa does not work out of the box. There are encoding issues. Therefore, you can do the following:

   # download
   pip download fever-drqa # or download directly from: https://pypi.org/project/fever-drqa/#files
   # extract
   tar -xvf fever-drqa-VERSION.tar.gz # where VERSION is the version you downloaded
   cd fever-drqa-VERSION

   In the extracted folder, you need to change setup.py:

   - with open('README.md') as f:
   + with open('README.md', encoding="utf-8") as f:
       readme = f.read()
   
   - with open('LICENSE') as f:
   + with open('LICENSE', encoding="utf-8") as f:
       license = f.read()
   
   - with open('requirements.txt') as f:
   + with open('requirements.txt', encoding="utf-8") as f:
       reqs = f.read()

   and also the drqa/__init__.py file:

   - from pathlib import PosixPath
   + from pathlib import Path
   
   # ...
   
   DATA_DIR = (
       os.getenv('DRQA_DATA') or
   -    os.path.join(PosixPath(__file__).absolute().parents[1].as_posix(), 'data')
   +    os.path.join(Path(__file__).absolute().parents[1], 'data')
   )

   Then, install it with pip:

   pip install .

   and clean up the files:

   # upon successful install, clean up the files
   cd ..
   rm -rf fever-drqa-VERSION
   rm -rf fever-drqa-VERSION.tar.gz

2. Download necessary models

   python -m spacy download en_core_web_sm

3. Download Wikipedia processed dump (kilt_knowledgesource.json) from KILT-github into data directory (Refer to their repository for citation details)

   mkdir data
   cd data
   wget http://dl.fbaipublicfiles.com/KILT/kilt_knowledgesource.json

4. Create the DB file from Wikipedia dump by running:

   PYTHONPATH=fever_athene python3 fever_athene/src/scripts/build_db_kilt.py data/kilt_knowledgesource.json data/kilt_db.db

   This script will create kilt_db.db into data directory.

5. Configure src/const.py file.

2. Run evaluation script

Running any of the scripts below will save corresponding metric results into a file named $GEN_TO_EVALUATE_NAME_results.jsonl ($GEN_TO_EVALUATE_NAME refers to the file containing generations that you are trying to evaluate).

Factuality Metric (Hallucinated NE Error, Entailment Ratio)

for PROMPT_TYPE in factual nonfactual
do
    GEN_TO_EVALUATE_NAME=${PROMPT_TYPE}-CUSTOM-GEN-NAME.jsonl
    PYTHONPATH=. python src/evaluate_v3_final.py --prompt_type ${PROMPT_TYPE} --gen_path ${GEN_TO_EVALUATE_NAME}
done

Repetition

for PROMPT_TYPE in factual nonfactual
do
    GEN_TO_EVALUATE_NAME=${PROMPT_TYPE}-CUSTOM-GEN-NAME.jsonl
    python src/repetition.py ${GEN_TO_EVALUATE_NAME}  --final
done

Diversity

1. First obtain multiple generation files from your LM with different seed. In our paper, we used 10 random seeds, but you can use your own choice of seed count. If you are evaluating greedy, there is NO NEED to generate multiple seed, because all seed will result in same generation. Simply use 1 generation file.

2. Then run the below script:

   GEN_DIR=directory-containing-multi-seed-generation-files
   
   FILE_TEMPLATE=shared-string-between-multiple-seed-generation
   python src/distinct_n.py --gen_dir ${GEN_DIR} --file_template ${FILE_TEMPLATE} --number_of_seeds 10

Illustration of FILE_TEMPLATE:

• Let's assume your generation files are named as follows: factual_gen_seed1.jsonl, nonfactual_gen_seed1.jsonl, factual_gen_seed2.jsonl, nonfactual_gen_seed2.jsonl,...
• Then, your FILE_TEMPLATE will be "gen_seed"

3. Replicating our work with Megatron-LM (Note: we used v3.0.2)

Factual Nucleus Decoding

Refer to this link for nucleus sampling implementation in Megatron-LM repository.

Sentence Completition Loss

Step 1. Prepare the training corpus:

  python preprocess_data_megatron_lm.py \
      --input $CORPUS_PATH \
      --output-prefix $OUTPUT_FILE_PREFIX \
      --vocab-file gpt2-vocab.json \
      --merge-file gpt2-merges.txt \
      --tokenizer-type GPT2BPETokenizer \
      --append-eod  --workers 20 \
      --mask_type $MASKING_CHOICE_FOR_SC_LOSS_PIVOT

Possible choice for $MASKING_CHOICE_FOR_SC_LOSS_PIVOT:

• v2_all_after_ROOT: ROOT Pivot
• v3_all_after_half: Half Pivot
• v5_RANDOM_Mask: Random pivot

Step 2: Modify the Megatron-LM code to incorporate SC-loss masking

We just need to modify get_batch() function from https://github.com/NVIDIA/Megatron-LM/blob/main/pretrain_gpt.py file with below code snippet:

def get_batch(data_iterator):
    """Generate a batch"""
    args = get_args()
    tokenizer = get_tokenizer()

    # Items and their type.
    # keys = ['text'] # <- original code
    keys = ['text', 'ne_mask'] # <- our code
    datatype = torch.int64

    # Broadcast data.
    if data_iterator is not None:
        data = next(data_iterator)
    else:
        data = None
    data_b = mpu.broadcast_data(keys, data, datatype)

    # Unpack.
    tokens_ = data_b['text'].long()
    labels = tokens_[:, 1:].contiguous()
    tokens = tokens_[:, :-1].contiguous()

    # Get the masks and postition ids.
    attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
        tokens,
        tokenizer.eod,
        args.reset_position_ids,
        args.reset_attention_mask,
        args.eod_mask_loss)

    # '''Below three lines are our code'''
    if 'ne_mask' in keys:
        loss_mask_ = data_b['ne_mask'].long()
        loss_mask = loss_mask_[:, :-1].contiguous()
        
    return tokens, labels, loss_mask, attention_mask, position_ids

Step 3: Use the provided script in Megtraon-LM repository (https://github.com/NVIDIA/Megatron-LM#gpt-pretraining) to continue training the Megatron-GPT with SC-loss

• Set DATA_PATH to the preprocessed files generated from Step 1.
• Note that: since publicly available Megatron-LM checkpoint (345M) is smaller than the models used in our paper, same performance won't be replicated.

4. Replicating our work with Hugginface (v4.20.1)

Please refer to this repository -> https://github.com/nayeon7lee/factuality_enhanced_lm_hf.