QA with Hugging Face pipeline in NLP - Model Metrics & Evaluation

Metrics & Evaluation - QA with Hugging Face pipeline

Which metric matters for QA with Hugging Face pipeline and WHY

For question answering (QA) tasks using Hugging Face pipelines, the key metrics are Exact Match (EM) and F1 score. Exact Match measures how often the model's answer exactly matches the correct answer. F1 score measures the overlap between the predicted and true answer words, balancing precision and recall. These metrics matter because QA answers can be short phrases or sentences, so exact matches are strict, while F1 allows partial credit for close answers.

Confusion matrix or equivalent visualization

QA tasks do not use a traditional confusion matrix because answers are text, not classes. Instead, evaluation compares predicted answers to true answers using token-level overlap.

True answer: "Paris"
Predicted answer: "Paris"
Exact Match: 1 (correct)

True answer: "Paris"
Predicted answer: "the city of Paris"
Exact Match: 0 (not exact)
F1 score: calculated from word overlap
    

Precision vs Recall tradeoff with concrete examples

In QA, precision means how many words in the predicted answer are correct, and recall means how many words from the true answer were found. For example:

• Predicted: "Paris"
  • True: "Paris"
    Precision = 1, Recall = 1 (perfect)
• Predicted: "the city"
  • True: "Paris"
    Precision = 0 (no correct words), Recall = 0 (missed true answer)
• Predicted: "city of Paris"
  • True: "Paris"
    Precision = 1/3, Recall = 1 (all true words found but extra words included)

Good QA models balance precision and recall to get high F1 scores.

What "good" vs "bad" metric values look like for QA

Good QA models have Exact Match scores above 70% and F1 scores above 80% on standard datasets. Bad models have low EM (below 40%) and low F1 (below 50%), meaning answers are often wrong or incomplete.

Example:

• Good: EM = 75%, F1 = 85% (answers mostly correct and complete)
• Bad: EM = 30%, F1 = 45% (answers often wrong or missing key info)

Common pitfalls in QA metrics

• Exact Match too strict: Small differences like punctuation or articles cause zero score even if answer is close.
• Ignoring partial credit: Relying only on EM misses partial correct answers; F1 helps here.
• Data leakage: Training on test questions inflates scores falsely.
• Overfitting: High training scores but low test scores mean model memorizes answers, not generalizes.
• Ambiguous questions: Multiple correct answers can confuse metric calculations.

Self-check question

Your QA model has 85% Exact Match but only 50% F1 score. Is it good? Why or why not?

Answer: This is unusual because F1 should be equal or higher than EM. A low F1 suggests the model's answers are often exact but very short or missing parts, or there may be an error in calculation. You should check the evaluation method and ensure answers are fully captured. Generally, both EM and F1 should be high for a good QA model.

Key Result

Exact Match and F1 score are key metrics for QA; they measure exact correctness and partial answer overlap respectively.

Practice

1. What does the Hugging Face QA pipeline do when given a question and a context?

easy

A. It translates the question into another language.

B. It summarizes the context without answering the question.

C. It finds the answer to the question from the given context.

D. It generates a new question based on the context.

Solution

1. Step 1: Understand the QA pipeline purpose

   The QA pipeline is designed to find answers from a given text based on a question.

2. Step 2: Match function to options

   Only It finds the answer to the question from the given context. describes finding an answer from the context, which is the pipeline's main job.

3. Final Answer:

   It finds the answer to the question from the given context. -> Option C

4. Quick Check:

   QA pipeline = find answer from context [OK]

Hint: QA pipeline = question + context -> answer [OK]

Common Mistakes:

• Confusing QA with translation or summarization
• Thinking it generates new questions
• Assuming it works without context

2. Which of the following is the correct way to create a QA pipeline using Hugging Face Transformers in Python?

easy

A. import pipeline from transformers
qa = pipeline('qa')

B. from transformers import QA
qa = QA('pipeline')

C. from transformers import question_answering
qa = question_answering()

D. from transformers import pipeline
qa = pipeline('question-answering')

Solution

1. Step 1: Recall correct import and pipeline creation

   The correct import is from transformers import pipeline, then call pipeline('question-answering').

2. Step 2: Check each option syntax

   Only from transformers import pipeline qa = pipeline('question-answering') matches the correct syntax and function call.

3. Final Answer:

   from transformers import pipeline qa = pipeline('question-answering') -> Option D

4. Quick Check:

   Correct import and pipeline call = from transformers import pipeline qa = pipeline('question-answering') [OK]

Hint: Use pipeline('question-answering') from transformers [OK]

Common Mistakes:

• Wrong import statement
• Incorrect pipeline argument
• Using non-existent classes or functions

3. What will be the output of this code snippet?

from transformers import pipeline
qa = pipeline('question-answering')
result = qa(question='Where is the Eiffel Tower?', context='The Eiffel Tower is in Paris.')
print(result['answer'])

medium

A. In Paris

B. Paris

C. The Eiffel Tower

D. Eiffel Tower

Solution

1. Step 1: Understand the question and context

   The question asks for the location of the Eiffel Tower, and the context states it is in Paris.

2. Step 2: Predict the pipeline answer output

   The pipeline extracts the answer span from the context, which is 'Paris'.

3. Final Answer:

   Paris -> Option B

4. Quick Check:

   Answer extracted = Paris [OK]

Hint: Answer is the location mentioned in context [OK]

Common Mistakes:

• Choosing the full phrase instead of the exact answer
• Confusing question with context text
• Expecting the pipeline to generate new text

4. Identify the error in this code snippet that uses the Hugging Face QA pipeline:

from transformers import pipeline
qa = pipeline('question-answering')
result = qa(question='Who wrote Hamlet?', text='Hamlet was written by Shakespeare.')
print(result['answer'])

medium

A. The argument 'text' should be 'context'.

B. The pipeline name should be 'qa' instead of 'question-answering'.

C. The print statement should use result.answer instead of result['answer'].

D. The import statement is incorrect.

Solution

1. Step 1: Check pipeline argument names

   The QA pipeline expects 'question' and 'context' as arguments, not 'text'.

2. Step 2: Verify other parts of the code

   Pipeline name and import are correct; accessing result['answer'] is valid.

3. Final Answer:

   The argument 'text' should be 'context'. -> Option A

4. Quick Check:

   Use 'context' argument for QA pipeline [OK]

Hint: Use 'context' not 'text' for QA input [OK]

Common Mistakes:

• Using 'text' instead of 'context'
• Changing pipeline name incorrectly
• Wrong result access syntax

5. You want to build a QA system that answers questions from multiple documents. Which approach using Hugging Face pipelines is best?

hard

A. Run the QA pipeline separately on each document and pick the answer with highest score.

B. Concatenate all documents into one string and run the QA pipeline once.

C. Use the QA pipeline only on the first document and ignore others.

D. Train a new model from scratch for multiple documents.

Solution

1. Step 1: Understand pipeline input limits

   QA pipelines work best on one context at a time; long concatenated text may reduce accuracy.

2. Step 2: Evaluate options for multiple documents

   Running QA on each document separately and selecting the best answer is effective and practical.

3. Final Answer:

   Run the QA pipeline separately on each document and pick the answer with highest score. -> Option A

4. Quick Check:

   Separate runs + best score = best multi-doc QA [OK]

Hint: Run QA on each doc, choose best answer [OK]

Common Mistakes:

• Concatenating all documents causing context overflow
• Ignoring documents except first
• Unnecessarily retraining models