Skip to content

Latest commit

 

History

History
89 lines (67 loc) · 3.87 KB

README.md

File metadata and controls

89 lines (67 loc) · 3.87 KB
Raw

Evaluating Question Answering with LLMs

Quick Links

Overview

Lexical matching is the standard evaluation method for open-domain question answering (QA), but it fails when plausible answers are not in the provided list. In this repo, we use open-source and proprietary LLMs for evaluation.

Requirements

The code needs Python 3.8+ (we tested it with Python 3.8).

To install from the repo:

pip install git+https://github.com/ehsk/QA-eval.git

To install from the source:

git clone [email protected]:ehsk/QA-eval.git
pip install -e .

Data

We worked on the Natural Questions-open (Lee et al., ACL 2019) test dataset that consists of 3,610 questions. We randomly sampled 301 questions for annotation.

Taken from here, we provide the answers generated by QA models along with the output of the four evaluation mechanisms in the data directory:

   data
    ├── model_outputs                                   # Answers generated by 12 open-domain QA models
    │   ├── NQ301_text-davinci-003_fewshot-n64.jsonl    # InstructGPT (few-shot)
    │   ├── NQ301_text-davinci-003_zeroshot.jsonl       # InstructGPT (zero-shot)
    │   ├── NQ_ANCE-plus_FiD.jsonl                      # ANCE+ & Fusion-In-Decoder
    │   └── ...
    ├── NQ301_BEM.tsv                                   # BEM predictions for all generated answers
    ├── NQ301_gpt-4.tsv                                 # GPT4-eval output for all generated answers
    ├── NQ301_human.tsv                                 # Human judgments for all generated answers
    └── NQ301_text-davinci-003.tsv                      # InstructGPT-eval output for all generated answers

The annotations can also be viewed online here.

Evaluation

The evaluation script takes a prediction file in a jsonl format as below and measures its performance with different metrics.

{"question": "who is under the mask of darth vader", "answer": ["Anakin Skywalker"], "prediction": "Anakin Skywalker"}
{"question": "which is the default file extension for an audio file in windows media player", "answer": ["Windows Playlist ( WPL )"], "prediction": "WMA"}

The following command computes only two lexical matching metrics: EM (Exact-Match accuracy) and macro-averaged F1.

python -m qaeval /path/to/prediction_file.jsonl

To evaluate using an LLM like InstructGPT-eval in the paper, the model name (text-davinci-003 or gpt-4) argument should be passed:

python -m qaeval /path/to/prediction_file.jsonl --model text-davinci-003

which calls OpenAI APIs. The environment variable OPENAI_API_KEY needs to be set first. Bear in mind that running this command will result in charges to your OpenAI account. We did not see a significant difference between GPT-4 and InstructGPT, so we recommend using the cheaper model (InstructGPT).

To evaluate using our provided annotated files including human judgment, you can simply run:

python -m qaeval /path/to/prediction_file.jsonl --annotation data/NQ301_human.tsv

The above command evaluates only 301 annotated questions and skips the rest in the prediction file.

Bugs:bug: or questions:question:

If you have any questions or encounter any problems, feel free to open an issue.

License

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