langchain/docs/use_cases/evaluation/question_answering.ipynb

Question Answering

This notebook covers how to evaluate generic question answering problems. This is a situation where you have an example containing a question and its corresponding ground truth answer, and you want to measure how well the language model does at answering those questions.

Setup

For demonstration purposes, we will just evaluate a simple question answering system that only evaluates the model's internal knowledge. Please see other notebooks for examples where it evaluates how the model does at question answering over data not present in what the model was trained on.

In [1]:

from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain.llms import OpenAI

In [2]:

prompt = PromptTemplate(template="Question: {question}\nAnswer:", input_variables=["question"])

In [3]:

llm = OpenAI(model_name="text-davinci-003", temperature=0)
chain = LLMChain(llm=llm, prompt=prompt)

Examples

For this purpose, we will just use two simple hardcoded examples, but see other notebooks for tips on how to get and/or generate these examples.

In [4]:

examples = [
    {
        "question": "Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?",
        "answer": "11"
    },
    {
        "question": 'Is the following sentence plausible? "Joao Moutinho caught the screen pass in the NFC championship."',
        "answer": "No"
    }
]

Predictions

We can now make and inspect the predictions for these questions.

In [5]:

predictions = chain.apply(examples)

In [6]:

predictions

Out [6]:

[{'text': ' 11 tennis balls'},
 {'text': ' No, this sentence is not plausible. Joao Moutinho is a professional soccer player, not an American football player, so it is not likely that he would be catching a screen pass in the NFC championship.'}]

Evaluation

We can see that if we tried to just do exact match on the answer answers (11 and No) they would not match what the language model answered. However, semantically the language model is correct in both cases. In order to account for this, we can use a language model itself to evaluate the answers.

In [7]:

from langchain.evaluation.qa import QAEvalChain

In [8]:

llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(examples, predictions, question_key="question", prediction_key="text")

In [9]:

for i, eg in enumerate(examples):
    print(f"Example {i}:")
    print("Question: " + eg['question'])
    print("Real Answer: " + eg['answer'])
    print("Predicted Answer: " + predictions[i]['text'])
    print("Predicted Grade: " + graded_outputs[i]['text'])
    print()

Example 0:
Question: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?
Real Answer: 11
Predicted Answer:  11 tennis balls
Predicted Grade:  CORRECT

Example 1:
Question: Is the following sentence plausible? "Joao Moutinho caught the screen pass in the NFC championship."
Real Answer: No
Predicted Answer:  No, this sentence is not plausible. Joao Moutinho is a professional soccer player, not an American football player, so it is not likely that he would be catching a screen pass in the NFC championship.
Predicted Grade:  CORRECT

Customize Prompt

You can also customize the prompt that is used. Here is an example prompting it using a score from 0 to 10. The custom prompt requires 3 input variables: "query", "answer" and "result". Where "query" is the question, "answer" is the ground truth answer, and "result" is the predicted answer.

In [ ]:

from langchain.prompts.prompt import PromptTemplate

_PROMPT_TEMPLATE = """You are an expert professor specialized in grading students' answers to questions.
You are grading the following question:
{query}
Here is the real answer:
{answer}
You are grading the following predicted answer:
{result}
What grade do you give from 0 to 10, where 0 is the lowest (very low similarity) and 10 is the highest (very high similarity)?
"""

PROMPT = PromptTemplate(input_variables=["query", "answer", "result"], template=_PROMPT_TEMPLATE)

In [ ]:

evalchain = QAEvalChain.from_llm(llm=llm,prompt=PROMPT)
evalchain.evaluate(examples, predictions, question_key="question", answer_key="answer", prediction_key="text")

Evaluation without Ground Truth

Its possible to evaluate question answering systems without ground truth. You would need a "context" input that reflects what the information the LLM uses to answer the question. This context can be obtained by any retreival system. Here's an example of how it works:

In [ ]:

context_examples = [
    {
        "question": "How old am I?",
        "context": "I am 30 years old. I live in New York and take the train to work everyday.",
    },
    {
        "question": 'Who won the NFC championship game in 2023?"',
        "context": "NFC Championship Game 2023: Philadelphia Eagles 31, San Francisco 49ers 7"
    }
]
QA_PROMPT = "Answer the question based on the  context\nContext:{context}\nQuestion:{question}\nAnswer:"
template = PromptTemplate(input_variables=["context", "question"], template=QA_PROMPT)
qa_chain = LLMChain(llm=llm, prompt=template)
predictions = qa_chain.apply(context_examples)

In [12]:

predictions

Out [12]:

[{'text': 'You are 30 years old.'},
 {'text': ' The Philadelphia Eagles won the NFC championship game in 2023.'}]

In [9]:

from langchain.evaluation.qa import ContextQAEvalChain
eval_chain = ContextQAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(context_examples, predictions, question_key="question", prediction_key="text")

In [13]:

graded_outputs

Out [13]:

[{'text': ' CORRECT'}, {'text': ' CORRECT'}]

Comparing to other evaluation metrics

We can compare the evaluation results we get to other common evaluation metrics. To do this, let's load some evaluation metrics from HuggingFace's evaluate package.

In [10]:

# Some data munging to get the examples in the right format
for i, eg in enumerate(examples):
    eg['id'] = str(i)
    eg['answers'] = {"text": [eg['answer']], "answer_start": [0]}
    predictions[i]['id'] = str(i)
    predictions[i]['prediction_text'] = predictions[i]['text']

for p in predictions:
    del p['text']

new_examples = examples.copy()
for eg in new_examples:
    del eg ['question']
    del eg['answer']

In [11]:

from evaluate import load
squad_metric = load("squad")
results = squad_metric.compute(
    references=new_examples,
    predictions=predictions,
)

In [12]:

results

Out [12]:

{'exact_match': 0.0, 'f1': 28.125}

In [ ]: