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:

📖 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

@misc{li2024pedantscheapeffectiveinterpretable,
      title={PEDANTS: Cheap but Effective and Interpretable Answer Equivalence}, 
      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},
      url={https://arxiv.org/abs/2402.11161}, 
}

📝 License

This project is licensed under the MIT License.

📬 Contact

For questions or comments, please contact: [email protected]