Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering

Quick Start

Installation

Make sure you have Python 3.7+ installed. It is also a good idea to use a virtual environment.

python3 -m venv instruct-qa-venv
source instruct-qa-venv/bin/activate

You can install the library via pip:

# Install the latest release
pip3 install instruct-qa

# Install the latest version from GitHub
pip3 install git+https://github.com/McGill-NLP/instruct-qa

For development, you can install it in editable mode with:

git clone https://github.com/McGill-NLP/instruct-qa
cd instruct-qa/
pip3 install -e .

Usage

Here is a simple example to get started. Using this library, use can easily leverage retrieval-augmented instruction-following models for question-answering in ~25 lines of code. The source file for this example is examples/get_started.py.

from instruct_qa.collections.utils import load_collection
from instruct_qa.retrieval.utils import load_retriever, load_index
from instruct_qa.prompt.utils import load_template
from instruct_qa.generation.utils import load_model
from instruct_qa.response_runner import ResponseRunner

collection = load_collection("dpr_wiki_collection")
index = load_index("dpr-nq-multi-hnsw")
retriever = load_retriever("facebook-dpr-question_encoder-multiset-base", index)
model = load_model("flan-t5-xxl")
prompt_template = load_template("qa")

queries = ["what is haleys comet"]

runner = ResponseRunner(
    model=model,
    retriever=retriever,
    document_collection=collection,
    prompt_template=prompt_template,
    queries=queries,
)

responses = runner()
print(responses[0]["response"])
# Halley's Comet Halley's Comet or Comet Halley, officially designated 1P/Halley, is a short-period comet visible from Earth every 75–76 years. Halley is the only known short-period comet that is regularly visible to the naked eye from Earth, and the only naked-eye comet that might appear twice in a human lifetime. Halley last appeared...

You can also check the input prompt given to the instruction-sollowing model that contains the instruction and the retrieved passages.

print(responses[0]["prompt"])
"""
Please answer the following question given the following passages:
- Title: Bill Haley
then known as Bill Haley's Saddlemen...

- Title: C/2016 R2 (PANSTARRS)
(CO) with a blue coma. The blue color...

...

Question: what is haleys comet
Answer:
"""

Detailed documentation of different modules of the library can be found here

Generating responses for entire datasets

Our library supports both question answering (QA) and conversational question answering (ConvQA) datasets. The following datasets are currently incorporated in the library

• Natural Questions (Open-domain)
• HotpotQA
• TopiOCQA

Here is an example to generate responses for Natural Questions using DPR retriever and Flan-T5 generator.

python experiments/question_answering.py \
--prompt_type qa \
--dataset_name natural_questions \
--document_collection_name dpr_wiki_collection \
--index_name dpr-nq-multi-hnsw \
--retriever_name facebook-dpr-question_encoder-multiset-base \
--batch_size 1 \
--model_name flan-t5-xxl \
--k 8

By default, a results directory is created within the repository that stores the model responses. The default directory location can be overidden by providing an additional command line argument --persistent_dir <OUTPUT_DIR> More examples are present in the examples directory.

Download model responses and human evaluation data

We release the model responses generated using the above commands for all three datasets. The scores reported in the paper are based on these responses. The responses can be downloaded with the following command:

python download_data.py --resource results

The responses are automatically unzipped and stored as JSON lines in the following directory structure:

results
├── {dataset_name}
│   ├── response
│   │   ├── {dataset}_{split}_c-{collection}_m-{model}_r-{retriever}_prompt-{prompt}_p-{top_p}_t-{temperature}_s-{seed}.jsonl

Currently, the following models are included:

• fid (Fusion-in-Decoder, separately fine-tuned on each dataset)
• gpt-3.5-turbo (GPT-3.5)
• alpaca-7b (Alpaca)
• llama-2-7b-chat (Llama-2)
• flan-t5-xxl (Flan-T5)

We also release the human annotations for correctness and faithfulness on a subset of responses for all datasets. The annotations can be downloaded with the following command:

python download_data.py --resource human_eval_annotations

The responses will be automatically unzipped in the following directory structure:

human_eval_annotations
├── correctness
│   ├── {dataset_name}
│   │   ├── {model}_human_eval_results.json
|
├── faithfulness
│   ├── {dataset_name}
│   │   ├── {model}_human_eval_results.json

Evaluating model responses (Coming soon!)

Documentation to evaluate model responses and add your own evaluation criterion coming soon! Stay tuned!

License

This work is licensed under the Apache 2 license. See LICENSE for details.

Citation

To cite this work, please use the following citation:

@article{adlakha2023evaluating,
      title={Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering}, 
      author={Vaibhav Adlakha and Parishad BehnamGhader and Xing Han Lu and Nicholas Meade and Siva Reddy},
      year={2023},
      journal={arXiv:2307.16877},
}

Contact

For queries and clarifications please contact vaibhav.adlakha (at) mila (dot) quebec