Questions
EDSL provides templates for many common question types, including multiple choice, checkbox, free text, numerical, linear scale and others. The Question class has subclasses for each of these types: QuestionMultipleChoice, QuestionCheckBox, QuestionFreeText, QuestionNumerical, QuestionLinearScale, etc., which have methods for validating answers and responses from language models.
Question type templates
A question is constructed by creating an instance of a question type class and passing the required fields. Questions are formatted as dictionaries with specific keys based on the question type.
Question types
The following question types are available:
QuestionMultipleChoice - multiple choice questions
QuestionCheckBox - checkbox questions
QuestionFreeText - free text questions
QuestionNumerical - numerical questions
QuestionLinearScale - linear scale questions
QuestionLikertFive - Likert scale questions
QuestionRank - ranked list questions
QuestionTopK - top-k list questions
QuestionYesNo - yes/no questions (multiple choice with fixed options)
QuestionList - list questions
QuestionBudget - budget allocation questions
QuestionExtract - information extraction questions
QuestionFunctional - functional questions
Required fields
All question types require a question_name and question_text. The question_name is a unique Pythonic identifier for a question (e.g., “favorite_color” but not “favorite color”). The question_text is the text of the question itself written as a string (e.g., “What is your favorite color?”). Question types other than free text require a question_options list of possible answer options. The question_options list can be a list of strings, integers or other types depending on the question type.
For example, to create a multiple choice question where the respondent must select one option from a list of colors, we import the QuestionMultipleChoice class and create an instance of it with the required fields:
from edsl import QuestionMultipleChoice
q = QuestionMultipleChoice(
question_name = "favorite_color",
question_text = "What is your favorite color?",
question_options = ["Red", "Blue", "Green", "Yellow"]
)
Optional fields for specific question types
min_selections and max_selections - Parameters that can be added to checkbox and rank questions to specify the minimum and maximum number of options that can be selected. For example, in a checkbox question where the respondent can select multiple options, we can specify that at least 2 options must be selected and at most 3 can be selected:
from edsl import QuestionCheckBox
q = QuestionCheckBox(
question_name = "favorite_days",
question_text = "What are your favorite days of the week?",
question_options = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"],
min_selections = 2,
max_selections = 3
)
min_value and max_value - Parameters that can be added to numerical questions to specify the minimum and maximum values that can be entered. For example, in a numerical question where the respondent must enter a number between 1 and 10:
from edsl import QuestionNumerical
q = QuestionNumerical(
question_name = "age",
question_text = "How old are you?",
min_value = 1,
max_value = 10
)
option_labels - A parameter that can be added to linear_scale questions to specify labels for the scale options. For example, in a linear scale question where the respondent must rate their agreement with a statement on a scale from 1 to 5:
from edsl import QuestionLinearScale
q = QuestionLinearScale(
question_name = "agree",
question_text = "Please indicate whether you agree with the following statement: I am only happy when it rains.",
question_options = [1, 2, 3, 4, 5],
option_labels = {1: "Strongly disagree", 5: "Strongly agree"}
)
num_selections - A parameter that can be added to rank questions to specify the number of options that must be ranked. For example, in a rank question where the respondent must rank their top 3 favorite foods:
from edsl import QuestionRank
q = QuestionRank(
question_name = "foods_rank",
question_text = "Rank your top 3 favorite foods.",
question_options = ["Pizza", "Pasta", "Salad", "Soup"],
num_selections = 3
)
answer_template - A parameter that can be added to extract questions to specify a template for the extracted information. For example, in an extract question where the respondent must extract information from a given text:
from edsl import QuestionExtract
q = QuestionExtract(
question_name = "course_schedule",
question_text = """This semester we are offering courses on calligraphy on Friday mornings.""",
answer_template = {"course_topic": "AI", "days": ["Monday", "Wednesday"]}
)
func - A parameter that can be added to functional questions to specify a function that generates the answer. For example, in a functional question where the answer is generated by a function:
from edsl import QuestionFunctional, ScenarioList, Scenario
import random
scenarios = ScenarioList(
[Scenario({"persona": p, "random": random.randint(0, 1000)}) for p in ["Magician", "Economist"]]
)
def my_function(scenario, agent_traits):
if scenario.get("persona") == "Magician":
return "Magicians never pick randomly!"
elif scenario.get("random") > 500:
return "Top half"
else:
return "Bottom half"
q = QuestionFunctional(
question_name = "evaluate",
func = my_function
)
results = q.by(scenarios).run()
results.select("persona", "random", "evaluate").print(format="rich")
Example results:
┏━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ scenario ┃ scenario ┃ answer ┃
┃ .persona ┃ .random ┃ .evaluate ┃
┡━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ Magician │ 301 │ Magicians never pick randomly! │
├───────────┼──────────┼────────────────────────────────┤
│ Economist │ 395 │ Bottom half │
└───────────┴──────────┴────────────────────────────────┘
Optional additional parameters
include_comment = False - A boolean value that can be added to any question type (other than free text) to exclude the comment field.
By default, a comment field is automatically added to all question types other than free text. It is a free text field that allows a model to provide any commentary on its response to the question, such as why a certain answer was chosen or how the model arrived at its answer. It can be useful for debugging unexpected responses and reducing the likelihood that a model fails to follow formatting instructions for the main response (e.g., just selecting an answer option) because it wants to be more verbose. It can also be helpful in constructing sequences of questions with context about prior responses (e.g., simulating a chain of thought). (See the survey section for more information about adding question memory to a survey.)
question_presentation - A string that can be added to any question type to specify how the question should be presented to the model. It can be used to provide additional context or instructions to the model about how to interpret the question.
answering_instructions - A string that can be added to any question type to specify how the model should answer the question. It can be used to provide additional context or instructions to the model about how to format its response.
permissive = False - A boolean value that can be added to any question type to specify whether the model should be allowed to provide an answer that violates the question constraints (e.g., selecting fewer or more than the allowed number of options in a checkbox question). (By default, permissive is set to False to enforce any question constraints.)
Creating a survey
We can combine multiple questions into a survey by passing them as a list to a Survey object:
from edsl import QuestionLinearScale, QuestionFreeText, QuestionNumerical, Survey
q1 = QuestionLinearScale(
question_name = "likely_to_vote",
question_text = "On a scale from 1 to 5, how likely are you to vote in the upcoming U.S. election?",
question_options = [1, 2, 3, 4, 5],
option_labels = {1: "Not at all likely", 5: "Very likely"}
)
q2 = QuestionFreeText(
question_name = "largest_us_city",
question_text = "What is the largest U.S. city?"
)
q3 = QuestionNumerical(
question_name = "us_pop",
question_text = "What was the U.S. population in 2020?"
)
survey = Survey(questions = [q1, q2, q3])
results = survey.run()
This allows us to administer multiple questions at once, either asynchronously (by default) or according to specified logic (e.g., skip or stop rules). To learn more about designing surveys with conditional logic, please see the Surveys section.
Simulating a response
We generate a response to a question by delivering it to a language model. This is done by calling the run method for the question:
results = q.run()
This will generate a Results object that contains a single Result representing the response to the question and information about the model used. If the model to be used has not been specified (as in the above example), the run method delivers the question to the default LLM (GPT 4). We can inspect the response and model used by calling the select and print methods on the components of the results that we want to display. For example, we can print just the answer to the question:
results.select("answer.favorite_primary_color").print(format="rich")
Output:
┏━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ answer ┃
┃ .favorite_primary_color ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ blue │
└─────────────────────────┘
Or to inspect the model:
results.select("model").print(format="rich")
Output:
┏━━━━━━━━━━━━━━━━━━━━┓
┃ model ┃
┃ .model ┃
┡━━━━━━━━━━━━━━━━━━━━┩
│ gpt-4-1106-preview │
└────────────────────┘
If questions have been combined in a survey, the run method is called directly on the survey instead:
results = survey.run()
results.select("answer.*").print(format="rich")
Output:
┏━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
┃ answer ┃ answer ┃ answer ┃
┃ .likely_to_vote ┃ .largest_us_city ┃ .us_pop ┃
┡━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
│ 4 │ The largest U.S. city by population is New York City. │ 331449281 │
└─────────────────┴───────────────────────────────────────────────────────┴───────────┘
For a survey, each Result represents a response for the set of survey questions. To learn more about analyzing results, please see the Results section.
Parameterizing a question
A question can also be constructed to take one or more parameters that are replaced with specified values either when the question is constructed or when the question is run. This allows us to easily create and administer multiple versions of a question at once.
Key steps:
Create a question text that takes a parameter in double braces:
from edsl.questions import QuestionFreeText
q = QuestionFreeText(
question_name = "favorite_item",
question_text = "What is your favorite {{ item }}?",
)
Create a dictionary for each value that will replace the parameter and store them in Scenario objects:
from edsl import ScenarioList, Scenario
scenarios = ScenarioList(
Scenario({"item": item}) for item in ["color", "food"]
)
To create multiple versions of the question when constructing a survey (i.e., before running it), pass the scenarios to the question loop method:
questions = q.loop(scenarios)
We can inspect the questions that have been created:
questions
Output:
[Question('free_text', question_name = """favorite_item_0""", question_text = """What is your favorite color?"""),
Question('free_text', question_name = """favorite_item_1""", question_text = """What is your favorite food?""")]
Note that a unique question_name has been automatically generated based on the parameter values. We can also specify that the paramater values be inserted in the question name (so long as they are Pythonic):
from edsl import QuestionFreeText, ScenarioList, Scenario
q = QuestionFreeText(
question_name = "favorite_{{ item }}",
question_text = "What is your favorite {{ item }}?",
)
scenarios = ScenarioList(
Scenario({"item": item}) for item in ["color", "food"]
)
questions = q.loop(scenarios)
Output:
[Question('free_text', question_name = """favorite_color""", question_text = """What is your favorite color?"""),
Question('free_text', question_name = """favorite_food""", question_text = """What is your favorite food?""")]
To run the questions, we pass them to a Survey and then call the run method, as before:
from edsl import Survey
survey = Survey(questions = questions)
results = survey.run()
Alternatively, we can pass the scenario or scenarios to the question with the by method when the question is run:
from edsl import QuestionFreeText, ScenarioList, Scenario
q = QuestionFreeText(
question_name = "favorite_item",
question_text = "What is your favorite {{ item }}?",
)
scenarios = ScenarioList(
Scenario({"item": item}) for item in ["color", "food"]
)
results = q.by(scenarios).run()
Each of the Results that are generated will include an individual Result for each version of the question that was answered.
To learn more about using scenarios, please see the Scenarios section.
Designing AI agents
A key feature of EDSL is the ability to design AI agents with personas and other traits for language models to use in responding to questions. The use of agents allows us to simulate survey results for target audiences at scale. This is done by creating Agent objects with dictionaries of desired traits and adding them to questions when they are run. For example, if we want a question answered by an AI agent representing a student we can create an Agent object with a relevant persona and attributes:
from edsl import Agent
agent = Agent(traits = {
"persona": "You are a student...", # can be an extended text
"age": 20, # individual trait values can be useful for analysis
"current_grade": "college sophomore"
})
To generate a response for the agent, we pass it to the by method when we run the question:
results = q.by(agent).run()
We can also generate responses for multiple agents at once by passing them as a list:
from edsl import AgentList, Agent
agents = AgentList(
Agent(traits = {"persona":p}) for p in ["Dog catcher", "Magician", "Spy"]
)
results = q.by(scenarios).by(agents).run()
The Results will contain a Result for each agent that answered the question. To learn more about designing agents, please see the Agents section.
Specifying language models
In the above examples we did not specify a language model for the question or survey, so the default model (GPT 4) was used. Similar to the way that we optionally passed scenarios to a question and added AI agents, we can also use the by method to specify one or more LLMs to use in generating results. This is done by creating Model objects for desired models and optionally specifying model parameters, such as temperature.
To check available models:
from edsl import Model
Model.available()
This will return a list of names of models that we can choose from.
We can also check the models for which we have already added API keys:
Model.check_models()
See instructions on storing API Keys for the models that you want to use, or activating Remote Inference to use the Expected Parrot server to access available models.
To specify models for a survey we first create Model objects:
from edsl import ModelList, Model
models = ModelList(
Model(m) for m in ['claude-3-opus-20240229', 'llama-2-70b-chat-hf']
)
Then we add them to a question or survey with the by method when running it:
results = q.by(models).run()
If scenarios and/or agents are also specified, each component is added in its own by call, chained together in any order, with the run method appended last:
results = q.by(scenarios).by(agents).by(models).run()
Note that multiple scenarios, agents and models are always passed as lists in the same by call.
Learn more about specifying question scenarios, agents and language models and their parameters in the respective sections:
Question type classes
QuestionMultipleChoice class
A subclass of the Question class for creating multiple choice questions where the response is a single option selected from a list of options. It specially requires a question_options list of strings for the options. Example usage:
from edsl.questions import QuestionMultipleChoice
q = QuestionMultipleChoice(
question_name = "color",
question_text = "What is your favorite color?",
question_options = ["Red", "Blue", "Green", "Yellow"]
)
An example can also created using the example method:
QuestionMultipleChoice.example()
QuestionCheckBox class
A subclass of the Question class for creating questions where the response is a list of one or more of the given options. It specially requires a question_options list of strings for the options. The minimum number of options that must be selected and the maximum number that may be selected can be specified when creating the question (parameters min_selections and max_selections). If not specified, the minimum number of options that must be selected is 1 and the maximum allowed is the number of question options provided. Example usage:
from edsl.questions import QuestionCheckBox
q = QuestionCheckBox(
question_name = "favorite_days",
question_text = "What are your 2 favorite days of the week?",
question_options = ["Monday", "Tuesday", "Wednesday",
"Thursday", "Friday", "Saturday", "Sunday"],
min_selections = 2, # optional
max_selections = 2 # optional
)
An example can also be created using the example method:
QuestionCheckBox.example()
- class edsl.questions.QuestionCheckBox.CheckBoxResponseValidator(response_model: type[BaseModel], exception_to_throw: Exception | None = None, override_answer: dict | None = None, **kwargs)[source]
Bases:
ResponseValidatorABC
- invalid_examples = [({'answer': [-1]}, {'question_options': ['Good', 'Great', 'OK', 'Bad']}, 'Answer code must be a non-negative integer'), ({'answer': 1}, {'question_options': ['Good', 'Great', 'OK', 'Bad']}, 'Answer code must be a list'), ({'answer': [1, 2, 3, 4]}, {'max_selections': 2, 'min_selections': 1, 'question_options': ['Good', 'Great', 'OK', 'Bad']}, 'Too many options selected')][source]
- class edsl.questions.QuestionCheckBox.QuestionCheckBox(question_name: str, question_text: str, question_options: list[str], min_selections: int | None = None, max_selections: int | None = None, include_comment: bool = True, use_code: bool = True, question_presentation: str | None = None, answering_instructions: str | None = None, permissive: bool = False)[source]
Bases:
QuestionBase
This question prompts the agent to select options from a list.
- __init__(question_name: str, question_text: str, question_options: list[str], min_selections: int | None = None, max_selections: int | None = None, include_comment: bool = True, use_code: bool = True, question_presentation: str | None = None, answering_instructions: str | None = None, permissive: bool = False)[source]
Instantiate a new QuestionCheckBox.
- Parameters:
question_name – The name of the question.
question_text – The text of the question.
question_options – The options the respondent should select from.
min_selections – The minimum number of options that must be selected.
max_selections – The maximum number of options that must be selected.
- classmethod example(exception_to_throw: Exception | None = None, override_answer: dict | None = None, *args, **kwargs) T [source]
- response_validator_class[source]
alias of
CheckBoxResponseValidator
- edsl.questions.QuestionCheckBox.create_checkbox_response_model(choices: list, min_selections: int | None = None, max_selections: int | None = None, permissive: bool = False)[source]
Dynamically create a CheckboxResponse model with a predefined list of choices.
- Parameters:
choices – A list of allowed values for the answer field.
include_comment – Whether to include a comment field in the model.
- Returns:
A new Pydantic model class.
QuestionFreeText class
A subclass of the Question class for creating free response questions. There are no specially required fields (only question_name and question_text). The response is a single string of text. Example usage:
from edsl.questions import QuestionFreeText
q = QuestionFreeText(
question_name = "food",
question_text = "What is your favorite food?"
)
An example can also be created using the example method:
QuestionFreeText.example()
- class edsl.questions.QuestionFreeText.FreeTextResponse(*, answer: str, generated_tokens: str | None = None)[source]
Bases:
BaseModel
Validator for free text response questions.
- model_computed_fields: ClassVar[dict[str, ComputedFieldInfo]] = {}[source]
A dictionary of computed field names and their corresponding ComputedFieldInfo objects.
- model_config: ClassVar[ConfigDict] = {}[source]
Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].
- model_fields: ClassVar[dict[str, FieldInfo]] = {'answer': FieldInfo(annotation=str, required=True), 'generated_tokens': FieldInfo(annotation=Union[str, NoneType], required=False, default=None)}[source]
Metadata about the fields defined on the model, mapping of field names to [FieldInfo][pydantic.fields.FieldInfo].
This replaces Model.__fields__ from Pydantic V1.
- class edsl.questions.QuestionFreeText.FreeTextResponseValidator(response_model: type[BaseModel], exception_to_throw: Exception | None = None, override_answer: dict | None = None, **kwargs)[source]
Bases:
ResponseValidatorABC
- class edsl.questions.QuestionFreeText.QuestionFreeText(question_name: str, question_text: str, answering_instructions: str | None = None, question_presentation: str | None = None)[source]
Bases:
QuestionBase
This question prompts the agent to respond with free text.
- __init__(question_name: str, question_text: str, answering_instructions: str | None = None, question_presentation: str | None = None)[source]
Instantiate a new QuestionFreeText.
- Parameters:
question_name – The name of the question.
question_text – The text of the question.
- classmethod example(exception_to_throw: Exception | None = None, override_answer: dict | None = None, *args, **kwargs) T [source]
- response_validator_class[source]
alias of
FreeTextResponseValidator
QuestionLinearScale class
A subclass of the QuestionMultipleChoice class for creating linear scale questions. It requires a question_options list of integers for the scale. The option_labels parameter can be used to specify labels for the scale options. Example usage:
from edsl.questions import QuestionLinearScale
q = QuestionLinearScale(
question_name = "studying",
question_text = """On a scale from 0 to 5, how much do you
enjoy studying? (0 = not at all, 5 = very much)""",
question_options = [0, 1, 2, 3, 4, 5], # integers
option_labels = {0: "Not at all", 5: "Very much"} # optional
)
An example can also be created using the example method:
QuestionLinearScale.example()
QuestionNumerical class
A subclass of the Question class for creating questions where the response is a numerical value. The minimum and maximum values of the answer can be specified using the min_value and max_value parameters. Example usage:
from edsl.questions import QuestionNumerical
q = QuestionNumerical(
question_name = "work_days",
question_text = "How many days a week do you normally work?",
min_value = 1, # optional
max_value = 7 # optional
)
An example can also be created using the example method:
QuestionNumerical.example()
QuestionLikertFive class
A subclass of the QuestionMultipleChoice class for creating questions where the answer is a response to a given statement on a 5-point Likert scale. (The scale does not need to be added as a parameter.) Example usage:
from edsl.questions import QuestionLikertFive
q = QuestionLikertFive(
question_name = "happy",
question_text = "I am only happy when it rains."
)
An example can also be created using the example method:
QuestionLikertFive.example()
QuestionRank class
A subclass of the Question class for creating questions where the response is a ranked list of options. It specially requires a question_options list of strings for the options. The number of options that must be selected can be optionally specified when creating the question. If not specified, all options are included (ranked) in the response. Example usage:
from edsl.questions import QuestionRank
q = QuestionRank(
question_name = "foods_rank",
question_text = "Rank the following foods.",
question_options = ["Pizza", "Pasta", "Salad", "Soup"],
num_selections = 2 # optional
)
An example can also be created using the example method:
QuestionRank.example()
Alternatively, QuestionTopK can be used to ask the respondent to select a specific number of options from a list. (See the next section for details.)
QuestionTopK class
A subclass of the QuestionMultipleChoice class for creating questions where the response is a list of ranked items. It specially requires a question_options list of strings for the options and the number of options that must be selected (num_selections). Example usage:
from edsl.questions import QuestionTopK
q = QuestionTopK(
question_name = "foods_rank",
question_text = "Select the best foods.",
question_options = ["Pizza", "Pasta", "Salad", "Soup"],
min_selections = 2,
max_selections = 2
)
An example can also be created using the example method:
QuestionTopK.example()
QuestionYesNo class
A subclass of the QuestionMultipleChoice class for creating multiple choice questions where the answer options are already specified: [‘Yes’, ‘No’]. Example usage:
from edsl.questions import QuestionYesNo
q = QuestionYesNo(
question_name = "student",
question_text = "Are you a student?"
)
An example can also be created using the example method:
QuestionYesNo.example()
QuestionList class
A subclass of the Question class for creating questions where the response is a list of strings. The maximum number of items in the list can be specified using the max_list_items parameter. Example usage:
q = QuestionList(
question_name = "activities",
question_text = "What activities do you enjoy most?",
max_list_items = 5 # optional
)
An example can also be created using the example method:
QuestionList.example()
QuestionBudget class
A subclass of the Question class for creating questions where the response is an allocation of a sum among a list of options in the form of a dictionary where the keys are the options and the values are the allocated amounts. It specially requires a question_options list of strings for the options and a budget_sum number for the total sum to be allocated. Example usage:
from edsl.questions import QuestionBudget
q = QuestionBudget(
question_name = "food_budget",
question_text = "How would you allocate $100?",
question_options = ["Pizza", "Ice cream", "Burgers", "Salad"],
budget_sum = 100
)
An example can also be created using the example method:
QuestionBudget.example()
QuestionExtract class
A subclass of the Question class for creating questions where the response is information extracted (or extrapolated) from a given text and formatted according to a specified template. Example usage:
from edsl.questions import QuestionExtract
q = QuestionExtract(
question_name = "course_schedule",
question_text = """This semester we are offering courses on
calligraphy on Friday mornings.""",
answer_template = {"course_topic": "AI", "days": ["Monday",
"Wednesday"]}
)
An example can also be created using the example method:
QuestionExtract.example()
- class edsl.questions.QuestionExtract.ExtractResponseValidator(response_model: type[BaseModel], exception_to_throw: Exception | None = None, override_answer: dict | None = None, **kwargs)[source]
Bases:
ResponseValidatorABC
- class edsl.questions.QuestionExtract.QuestionExtract(question_text: str, answer_template: dict[str, Any], question_name: str, answering_instructions: str = None, question_presentation: str = None)[source]
Bases:
QuestionBase
This question prompts the agent to extract information from a string and return it in a given template.
- __init__(question_text: str, answer_template: dict[str, Any], question_name: str, answering_instructions: str = None, question_presentation: str = None)[source]
Initialize the question.
- Parameters:
question_name – The name of the question.
question_text – The text of the question.
question_options – The options the respondent should select from.
answer_template – The template for the answer.
- classmethod example(exception_to_throw: Exception | None = None, override_answer: dict | None = None, *args, **kwargs) T [source]
- response_validator_class[source]
alias of
ExtractResponseValidator
QuestionFunctional class
A subclass of the Question class for creating questions where the response is generated by a function instead of a lanugage model. The question type is not intended to be used directly in a survey, but rather to generate responses for other questions. This can be useful where a model is not needed for part of a survey, for questions that require some kind of initial computation, or for questions that are the result of a multi-step process. The question type lets us define a function func that takes in a scenario and (optional) agent traits and returns an answer.
Example usage:
Say we have some survey results where we asked some agents to pick a random number:
from edsl.questions import QuestionNumerical
from edsl import Agent
q_random = QuestionNumerical(
question_name = "random",
question_text = "Choose a random number between 1 and 1000."
)
agents = [Agent({"persona":p}) for p in ["Dog catcher", "Magician", "Spy"]]
results = q_random.by(agents).run()
results.select("persona", "random").print(format="rich")
The results are:
┏━━━━━━━━━━━━━┳━━━━━━━━━┓
┃ agent ┃ answer ┃
┃ .persona ┃ .random ┃
┡━━━━━━━━━━━━━╇━━━━━━━━━┩
│ Dog catcher │ 472 │
├─────────────┼─────────┤
│ Magician │ 537 │
├─────────────┼─────────┤
│ Spy │ 528 │
└─────────────┴─────────┘
We can use QuestionFunctional to evaluate the responses using a function instead of calling the language model to answer another question. The responses are passed to the function as scenarios, and then the function is passed to the QuestionFunctional object:
from edsl.questions import QuestionFunctional
def my_function(scenario, agent_traits):
if scenario.get("persona") == "Magician":
return "Magicians never pick randomly!"
elif scenario.get("random") > 500:
return "Top half"
else:
return "Bottom half"
q_evaluate = QuestionFunctional(
question_name = "evaluate",
func = my_function
)
Next we turn the responses into scenarios for the function:
scenarios = results.select("persona", "random").to_scenarios()
scenarios
We can inspect the scenarios:
[Scenario({'persona': 'Dog catcher', 'random': 472}),
Scenario({'persona': 'Magician', 'random': 537}),
Scenario({'persona': 'Spy', 'random': 528})]
Finally, we run the function with the scenarios:
results = q_evaluate.by(scenarios).run()
results.select("persona", "random", "evaluate").print(format="rich")
The results are:
┏━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ scenario ┃ scenario ┃ answer ┃
┃ .persona ┃ .random ┃ .evaluate ┃
┡━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ Dog catcher │ 472 │ Bottom half │
├─────────────┼──────────┼────────────────────────────────┤
│ Magician │ 537 │ Magicians never pick randomly! │
├─────────────┼──────────┼────────────────────────────────┤
│ Spy │ 528 │ Top half │
└─────────────┴──────────┴────────────────────────────────┘
Another example of QuestionFunctional can be seen in the following notebook, where we give agents different instructions for generating random numbers and then use a function to identify whether the responses are identical.
Example notebook: Simulating randomness
- class edsl.questions.QuestionFunctional.QuestionFunctional(question_name: str, func: Callable | None = None, question_text: str | None = 'Functional question', requires_loop: bool | None = False, function_source_code: str | None = None, function_name: str | None = None)[source]
Bases:
QuestionBase
A special type of question that is not answered by an LLM.
>>> from edsl import Scenario, Agent
# Create an instance of QuestionFunctional with the new function >>> question = QuestionFunctional.example()
# Activate and test the function >>> question.activate() >>> scenario = Scenario({“numbers”: [1, 2, 3, 4, 5]}) >>> agent = Agent(traits={“multiplier”: 10}) >>> results = question.by(scenario).by(agent).run() >>> results.select(“answer.*”).to_list()[0] == 150 True
# Serialize the question to a dictionary
>>> from edsl.questions.QuestionBase import QuestionBase >>> new_question = QuestionBase.from_dict(question.to_dict()) >>> results = new_question.by(scenario).by(agent).run() >>> results.select("answer.*").to_list()[0] == 150 True
- __init__(question_name: str, func: Callable | None = None, question_text: str | None = 'Functional question', requires_loop: bool | None = False, function_source_code: str | None = None, function_name: str | None = None)[source]
Other classes & methods
Settings for the questions module.