--- inference: false license: mit language: - en metrics: - exact_match - f1 - bertscore pipeline_tag: text-classification --- # QA-Evaluation-Metrics [![PyPI version qa-metrics](https://img.shields.io/pypi/v/qa-metrics.svg)](https://pypi.org/project/qa-metrics/) [![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/17b7vrZqH0Yun2AJaOXydYZxr3cw20Ga6?usp=sharing) QA-Evaluation-Metrics is a fast and lightweight Python package for evaluating question-answering models and prompting of black-box and open-source large language models. It provides various basic and efficient metrics to assess the performance of QA models. ### Updates - Uopdated to version 0.2.8 - Supports prompting OPENAI GPT-series models and Claude Series models now. (Assuimg OPENAI version > 1.0) - Supports prompting various open source models such as LLaMA-2-70B-chat, LLaVA-1.5 etc by calling API from [deepinfra](https://deepinfra.com/models). ## Installation * Python version >= 3.6 * openai version >= 1.0 To install the package, run the following command: ```bash pip install qa-metrics ``` ## Usage/Logistics The python package currently provides six QA evaluation methods. - Given a set of gold answers, a candidate answer to be evaluated, and a question (if applicable), the evaluation returns True if the candidate answer matches any one of the gold answer, False otherwise. - Different evaluation methods have distinct strictness of evaluating the correctness of a candidate answer. Some have higher correlation with human judgments than others. - Normalized Exact Match and Question/Answer type Evaluation are the most efficient method. They are suitable for short-form QA datasets such as NQ-OPEN, Hotpot QA, TriviaQA, SQuAD, etc. - Question/Answer Type Evaluation and Transformer Neural evaluations are cost free and suitable for short-form and longer-form QA datasets. They have higher correlation with human judgments than exact match and F1 score when the length of the gold and candidate answers become long. - Black-box LLM evaluations are closest to human evaluations, and they are not cost-free. ## Normalized Exact Match #### `em_match` Returns a boolean indicating whether there are any exact normalized matches between gold and candidate answers. **Parameters** - `reference_answer` (list of str): A list of gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. **Returns** - `boolean`: A boolean True/False signifying matches between reference or candidate answers. ```python from qa_metrics.em import em_match reference_answer = ["The Frog Prince", "The Princess and the Frog"] candidate_answer = "The movie \"The Princess and the Frog\" is loosely based off the Brother Grimm's \"Iron Henry\"" match_result = em_match(reference_answer, candidate_answer) print("Exact Match: ", match_result) ''' Exact Match: False ''' ``` ## F1 Score #### `f1_score_with_precision_recall` Calculates F1 score, precision, and recall between a reference and a candidate answer. **Parameters** - `reference_answer` (str): A gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. **Returns** - `dictionary`: A dictionary containing the F1 score, precision, and recall between a gold and candidate answer. ```python from qa_metrics.f1 import f1_match,f1_score_with_precision_recall f1_stats = f1_score_with_precision_recall(reference_answer[0], candidate_answer) print("F1 stats: ", f1_stats) ''' F1 stats: {'f1': 0.25, 'precision': 0.6666666666666666, 'recall': 0.15384615384615385} ''' match_result = f1_match(reference_answer, candidate_answer, threshold=0.5) print("F1 Match: ", match_result) ''' F1 Match: False ''' ``` ## Transformer Neural Evaluation Our fine-tuned BERT model is on 🤗 [Huggingface](https://huggingface.co/Zongxia/answer_equivalence_bert?text=The+goal+of+life+is+%5BMASK%5D.). Our Package also supports downloading and matching directly. [distilroberta](https://huggingface.co/Zongxia/answer_equivalence_distilroberta), [distilbert](https://huggingface.co/Zongxia/answer_equivalence_distilbert), [roberta](https://huggingface.co/Zongxia/answer_equivalence_roberta), and [roberta-large](https://huggingface.co/Zongxia/answer_equivalence_roberta-large) are also supported now! 🔥🔥🔥 #### `transformer_match` Returns True if the candidate answer is a match of any of the gold answers. **Parameters** - `reference_answer` (list of str): A list of gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. - `question` (str): The question for which the answers are being evaluated. **Returns** - `boolean`: A boolean True/False signifying matches between reference or candidate answers. ```python from qa_metrics.transformerMatcher import TransformerMatcher question = "Which movie is loosley based off the Brother Grimm's Iron Henry?" tm = TransformerMatcher("roberta-large") scores = tm.get_scores(reference_answer, candidate_answer, question) match_result = tm.transformer_match(reference_answer, candidate_answer, question) print("Score: %s; bert Match: %s" % (scores, match_result)) ''' Score: {'The Frog Prince': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.6934309}, 'The Princess and the Frog': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.7400551}}; TM Match: True ''' ``` ## Efficient and Robust Question/Answer Type Evaluation #### 1. `get_highest_score` Returns the gold answer and candidate answer pair that has the highest matching score. This function is useful for evaluating the closest match to a given candidate response based on a list of reference answers. **Parameters** - `reference_answer` (list of str): A list of gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. - `question` (str): The question for which the answers are being evaluated. **Returns** - `dictionary`: A dictionary containing the gold answer and candidate answer that have the highest matching score. #### 2. `get_scores` Returns all the gold answer and candidate answer pairs' matching scores. **Parameters** - `reference_answer` (list of str): A list of gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. - `question` (str): The question for which the answers are being evaluated. **Returns** - `dictionary`: A dictionary containing gold answers and the candidate answer's matching score. #### 3. `evaluate` Returns True if the candidate answer is a match of any of the gold answers. **Parameters** - `reference_answer` (list of str): A list of gold (correct) answers to the question. - `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated. - `question` (str): The question for which the answers are being evaluated. **Returns** - `boolean`: A boolean True/False signifying matches between reference or candidate answers. ```python from qa_metrics.pedant import PEDANT question = "Which movie is loosley based off the Brother Grimm's Iron Henry?" pedant = PEDANT() scores = pedant.get_scores(reference_answer, candidate_answer, question) max_pair, highest_scores = pedant.get_highest_score(reference_answer, candidate_answer, question) match_result = pedant.evaluate(reference_answer, candidate_answer, question) print("Max Pair: %s; Highest Score: %s" % (max_pair, highest_scores)) print("Score: %s; PANDA Match: %s" % (scores, match_result)) ''' Max Pair: ('the princess and the frog', 'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"'); Highest Score: 0.854451712151719 Score: {'the frog prince': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.7131625951317375}, 'the princess and the frog': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.854451712151719}}; PANDA Match: True ''' ``` ```python print(pedant.get_score(reference_answer[1], candidate_answer, question)) ''' 0.7122460127464126 ''' ``` ## Prompting LLM For Evaluation Note: The prompting function can be used for any prompting purposes. ###### OpenAI ```python from qa_metrics.prompt_llm import CloseLLM model = CloseLLM() model.set_openai_api_key(YOUR_OPENAI_KEY) prompt = 'question: What is the Capital of France?\nreference: Paris\ncandidate: The capital is Paris\nIs the candidate answer correct based on the question and reference answer? Please only output correct or incorrect.' model.prompt_gpt(prompt=prompt, model_engine='gpt-3.5-turbo', temperature=0.1, max_tokens=10) ''' 'correct' ''' ``` ###### Anthropic ```python model = CloseLLM() model.set_anthropic_api_key(YOUR_Anthropic_KEY) model.prompt_claude(prompt=prompt, model_engine='claude-v1', anthropic_version="2023-06-01", max_tokens_to_sample=100, temperature=0.7) ''' 'correct' ''' ``` ###### deepinfra (See below for descriptions of more models) ```python from qa_metrics.prompt_open_llm import OpenLLM model = OpenLLM() model.set_deepinfra_key(YOUR_DEEPINFRA_KEY) model.prompt(message=prompt, model_engine='mistralai/Mixtral-8x7B-Instruct-v0.1', temperature=0.1, max_tokens=10) ''' 'correct' ''' ``` If you find this repo avialable, please cite our paper: ```bibtex @misc{li2024panda, title={PANDA (Pedantic ANswer-correctness Determination and Adjudication):Improving Automatic Evaluation for Question Answering and Text Generation}, author={Zongxia Li and Ishani Mondal and Yijun Liang and Huy Nghiem and Jordan Lee Boyd-Graber}, year={2024}, eprint={2402.11161}, archivePrefix={arXiv}, primaryClass={cs.CL} } ``` ## Updates - [01/24/24] 🔥 The full paper is uploaded and can be accessed [here](https://arxiv.org/abs/2402.11161). The dataset is expanded and leaderboard is updated. - Our Training Dataset is adapted and augmented from [Bulian et al](https://github.com/google-research-datasets/answer-equivalence-dataset). Our [dataset repo](https://github.com/zli12321/Answer_Equivalence_Dataset.git) includes the augmented training set and QA evaluation testing sets discussed in our paper. - Now our model supports [distilroberta](https://huggingface.co/Zongxia/answer_equivalence_distilroberta), [distilbert](https://huggingface.co/Zongxia/answer_equivalence_distilbert), a smaller and more robust matching model than Bert! ## License This project is licensed under the [MIT License](LICENSE.md) - see the LICENSE file for details. ## Contact For any additional questions or comments, please contact [zli12321@umd.edu].