QA-Evaluation-Metrics 📊

A fast and lightweight Python package for evaluating question-answering models and prompting of black-box and open-source large language models.

pip install qa-metrics is all you need!

🤗 Huggingface Model and Dataset

Release Updates

• Version 0.2.42 Released! (06/20/2025)

  • RewardBert (ModerBert base) supports batch scores prediction to speed up prediction for RL training.

• Version 0.2.35 Released! (06/18/2025)

  • RewardBert (ModerBert base) trained to evaluate both short-form and long-form generations.
  • RewardBert outputs a likert scale between 1-5 or normalized score between 0-1.
  • Turn off nltk download verbose logs.

• Version 0.2.30 Released!

  • Enhanced PEDANTS with multi-pipeline support and improved edge case handling
  • Introduced trained tiny-bert for QA evaluation (18MB model size)
  • Added direct Huggingface model download support for TransformerMatcher

🚀 Quick Start

Table of Contents

• 1. RewardBert
• 2. Normalized Exact Match
• 2. Token F1 Score
• 3. PEDANTS
• 4. Finetuned Neural Matching
• 5. Prompting LLM

Prerequisites

• Python >= 3.6
• openai >= 1.0

Installation

pip install qa-metrics

💡 Features

Our package offers six QA evaluation methods with varying strengths:

Method	Best For	Cost	Correlation with Human Judgment
RewardBert	General Text Generations	Free	Very High
Normalized Exact Match	Short-form QA (NQ-OPEN, HotpotQA, etc.)	Free	Good
PEDANTS	Both short & medium-form QA	Free	Very High
Neural Evaluation	Both short & long-form QA	Free	High
Open Source LLM Evaluation	All QA types	Free	High
Black-box LLM Evaluation	All QA types	Paid	Highest

📖 Documentation

1. RewardBert

Method: compute_score

Parameters

• reference_answer (str): gold (correct) answer to the question
• candidate_answer (str): The answer provided by a candidate that needs to be evaluated

Returns

• tuple: A tuple of normalized and raw scores.

from qa_metrics.RewardBert import RewardBert

rb = RewardBert(device='cuda')
reference_answer = "The Frog Prince"
candidate_answer = "The movie \"The Princess and the Frog\" is loosely based off the Brother Grimm's \"Iron Henry\""
rb.compute_score(reference_answer, candidate_answer)
# (0.29113227128982544, 2.1645290851593018)

Method: compute_batch_scores

Parameters

• reference_answers (list of str): A list of gold (correct) answers to the question
• candidate_answer (list of str): A list of answers provided by a candidate that needs to be evaluated
• batch_size (int): batch size to predict (default 1)

Returns

• tuple: A tuple of a list of normalized and raw scores.

from qa_metrics.RewardBert import RewardBert

rb = RewardBert(device='cuda')
reference_answer = ["The Frog Prince"]
candidate_answer = ["The movie \"The Princess and the Frog\" is loosely based off the Brother Grimm's \"Iron Henry\""]
rb.compute_batch_scores(reference_answer, candidate_answer, batch_size=1)
# ([0.29113227128982544], [2.1645290851593018])

2. Normalized Exact Match

Method: em_match

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: True if there are any exact normalized matches between gold and candidate answers

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)

3. F1 Score

Method: f1_score_with_precision_recall

Parameters

• reference_answer (str): A gold (correct) answer to the question
• candidate_answer (str): The answer provided by a candidate that needs to be evaluated

Returns

• dictionary: Contains the F1 score, precision, and recall between a gold and candidate answer

Method: f1_match

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• threshold (float): F1 score threshold for considering a match (default: 0.5)

Returns

• boolean: True if F1 score exceeds threshold for any gold answer

from qa_metrics.f1 import f1_match, f1_score_with_precision_recall

f1_stats = f1_score_with_precision_recall(reference_answer[0], candidate_answer)
match_result = f1_match(reference_answer, candidate_answer, threshold=0.5)

4. PEDANTS

Method: get_score

Parameters

• reference_answer (str): A Gold answer
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• float: The similarity score between two strings (0 to 1)

Method: get_highest_score

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• dictionary: Contains the gold answer and candidate answer pair with highest matching score

Method: get_scores

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• dictionary: Contains matching scores for all gold answer and candidate answer pairs

Method: evaluate

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• boolean: True if candidate answer matches any gold answer

Method: get_question_type

Parameters

• reference_answer (list of str): List of gold answers
• question (str): The question being evaluated

Returns

• list: The type of the question (what, who, when, how, why, which, where)

Method: get_judgement_type

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• list: A list revised rules applicable to judge answer correctness

from qa_metrics.pedant import PEDANT

pedant = PEDANT()
scores = pedant.get_scores(reference_answer, candidate_answer, question)
match_result = pedant.evaluate(reference_answer, candidate_answer, question)

5. Transformer Neural Evaluation

Method: get_score

Parameters

• reference_answer (str): A Gold answer
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• float: The similarity score between two strings (0 to 1)

Method: get_highest_score

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• dictionary: Contains the gold answer and candidate answer pair with highest matching score

Method: get_scores

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• dictionary: Contains matching scores for all gold answer and candidate answer pairs

Method: transformer_match

Parameters

• reference_answer (list of str): List of gold answers
• candidate_answer (str): Candidate answer to evaluate
• question (str): The question being evaluated

Returns

• boolean: True if transformer model considers candidate answer equivalent to any gold answer

from qa_metrics.transformerMatcher import TransformerMatcher

### supports zli12321/roberta-large-qa-evaluator, `zli12321/answer_equivalence_bert`, `zli12321/answer_equivalence_distilbert`, `zli12321/answer_equivalence_roberta`, `zli12321/answer_equivalence_distilroberta`
tm = TransformerMatcher("zli12321/answer_equivalence_tiny_bert")
match_result = tm.transformer_match(reference_answer, candidate_answer, question)

6. LLM Integration

Method: prompt_gpt

Parameters

• prompt (str): The input prompt text
• model_engine (str): OpenAI model to use (e.g., 'gpt-3.5-turbo')
• temperature (float): Controls randomness (0-1)
• max_tokens (int): Maximum tokens in response

from qa_metrics.prompt_llm import CloseLLM

model = CloseLLM()
model.set_openai_api_key(YOUR_OPENAI_KEY)
result = model.prompt_gpt(prompt=prompt, model_engine='gpt-3.5-turbo')

Method: prompt_claude

Parameters

• prompt (str): The input prompt text
• model_engine (str): Claude model to use
• anthropic_version (str): API version
• max_tokens_to_sample (int): Maximum tokens in response
• temperature (float): Controls randomness (0-1)

model = CloseLLM()
model.set_anthropic_api_key(YOUR_ANTHROPIC_KEY)
result = model.prompt_claude(prompt=prompt, model_engine='claude-v1')

Method: prompt

Parameters

• message (str): The input message text
• model_engine (str): Model to use
• temperature (float): Controls randomness (0-1)
• max_tokens (int): Maximum tokens in response

from qa_metrics.prompt_open_llm import OpenLLM

model = OpenLLM()
model.set_deepinfra_key(YOUR_DEEPINFRA_KEY)
result = model.prompt(message=prompt, model_engine='mistralai/Mixtral-8x7B-Instruct-v0.1')

🤗 Model Hub

Our fine-tuned models are available on Huggingface:

• BERT
• DistilRoBERTa
• DistilBERT
• RoBERTa
• Tiny-BERT
• RoBERTa-Large

📚 Resources

• Full Paper
• Dataset Repository
• Supported Models on Deepinfra

📄 Citation

@inproceedings{li-etal-2024-pedants,
    title = "{PEDANTS}: Cheap but Effective and Interpretable Answer Equivalence",
    author = "Li, Zongxia  and
      Mondal, Ishani  and
      Nghiem, Huy  and
      Liang, Yijun  and
      Boyd-Graber, Jordan Lee",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.548/",
    doi = "10.18653/v1/2024.findings-emnlp.548",
    pages = "9373--9398",
    abstract = "Question answering (QA) can only make progress if we know if an answer is correct, but current answer correctness (AC) metrics struggle with verbose, free-form answers from large language models (LLMs). There are two challenges with current short-form QA evaluations: a lack of diverse styles of evaluation data and an over-reliance on expensive and slow LLMs. LLM-based scorers correlate better with humans, but this expensive task has only been tested on limited QA datasets. We rectify these issues by providing rubrics and datasets for evaluating machine QA adopted from the Trivia community. We also propose an efficient, and interpretable QA evaluation that is more stable than an exact match and neural methods (BERTScore)."
}

📝 License

This project is licensed under the MIT License.

📬 Contact

For questions or comments, please contact: zli12321@umd.edu