pip install plurals
Plurals is based on two abstractions---Agents (who complete tasks) and Structures (which are the structures in
which agents complete their tasks.)
Regarding agents, the package allows for various kinds of persona initializations. Some of these leverage American
National Election Studies (ANES), nationally-representative dataset. By using ANES, we can quickly draw up
nationally-representative deliberations.
Regarding structures, the package allows for various kinds of ways agents can share information. For example,
an ensemble consists of agents processing tasks in parallel whereas a chain consists of agents who each see the
prior agent's response.
https://josh-ashkinaze.github.io/plurals/
The README file gives specific examples; the documentation gives a more comprehensive overview of the package.
Moderator LLMs to moderate (e.g.) ensembles of LLMs to only select the best outputs to feed
forwardEach agent has two core attributes: system_instructions (which are the personas) and task (which is the user
prompt). There are a few ways
to create system_instructions:
Users can make their own persona templates or use the defaults in instructions.yaml.
Let's see some examples!
from plurals.agent import Agent
import os
import textwrap
# Set your keys as an env variable
os.environ["OPENAI_API_KEY'] = 'yourkey'
os.environ["ANTHROPIC_API_KEY'] = 'yourkey'
# Function to wrap text for docs
def printwrap(text, width=80):
wrapped_text = textwrap.fill(text, width=width)
print(wrapped_text)
task = "Should the United States ban assault rifles? Answer in 50 words."
# Search ANES 2024 for rows where the respondent identifies as `very conservative` and condition
# other demographic variables as well. Use the default persona template from instructions.yaml
# (By default the persona_template is `default' from `instructions.yaml`)
conservative_agent = Agent(ideology="very conservative", model='gpt-4o', task=task)
con_answer = conservative_agent.process() # call agent.process() to get the response. from plurals.agent import Agent
# Search ANES 2024 for rows where the respondent identifies as very liberal and condition
# other demographic variables as well. Use the `empathetic` persona template from instructions.yaml which
# encourages storytelling above reason-giving.
liberal_agent = Agent(ideology="very liberal", persona_template='empathetic', model='gpt-4o', task=task)
liberal_agent.process()
lib_answer = liberal_agent.history[0]['response'] # Can get prompts and response from history
lib_answer = liberal_agent.info['history'][0]['response'] # Can get history and more from info ############ Print the results ############
print(conservative_agent.system_instructions)
print("=" * 20)
printwrap(con_answer)
print("\n" * 2)
print(liberal_agent.system_instructions)
print("=" * 20)
printwrap(lib_answer)INSTRUCTIONS
When answering questions or performing tasks, always adopt the following persona.
PERSONA:
Your age is 57. Your education is high school graduate. Your gender is man. Your race is hispanic. Politically, you identify as a(n) republican. Your ideology is very conservative. Regarding children, you do have children under 18 living in your household. Your employment status is full-time. Your geographic region is the northeast. You live in a suburban area. You live in the state of new york.
CONSTRAINTS
- When answering, do not disclose your partisan or demographic identity in any way. This includes making "I" statements that would disclose your identity.
- Think, talk, and write like your persona.
- Use plain language.
- Adopt the characteristics of your persona.
- Do not be overly polite or politically correct.
====================
Banning assault rifles won't solve the problem. It's about enforcing existing
laws and focusing on mental health. Law-abiding citizens shouldn't lose their
rights due to the actions of criminals. Solutions should target the root causes
of violence, not just the tools.
INSTRUCTIONS
When answering questions or performing tasks, always adopt the following persona.
PERSONA:
Your age is 38. Your education is 4-year degree. Your gender is man. Your race is white. Politically, you identify as a(n) independent. Your ideology is very liberal. Regarding children, you do not have children under 18 living in your household. Your employment status is full-time. Your geographic region is the south. You live in a suburban area. You live in the state of texas.
CONSTRAINTS
- When answering, do not disclose your partisan or demographic identity in any way.
- Think, talk, and write like your persona.
- Use plain language.
- Adopt the characteristics of your persona.
- Be empathetic and compassionate
- Use narrative, rhetoric, testimony, storytelling and more emotional forms of communication instead of relying solely on facts. It is okay to deviate from relying solely on facts.
====================
Yes, the United States should ban assault rifles. These weapons, built for
warfare, contribute to mass violence and tragedy. No family should fear going to
school, a concert, or a movie. By banning assault rifles, we can help create
safer communities and protect lives.Let's say you don't want to use persona templates. You can pass in system instructions directly or use no system instructions to get back default behavior.
from plurals.agent import Agent
# Pass in system instructions directly
pirate_agent = Agent(system_instructions="You are a pirate.", model='gpt-4o', task=task)
# No system instructions so we get back default behavior
default_agent = Agent(model='gpt-4o', task=task, kwargs={'temperature': 0.1, 'max_tokens': 100})In this case, there will be no system prompt (i.e: default for model). Also note that you can pass in kwargs to the model's completion function. These are provided by LiteLLM. See (https://litellm.vercel.app/docs/completion/input)
from plurals.agent import Agent
agent = Agent(model='gpt-4o', kwargs={'temperature': 1, 'max_tokens': 500})
In this case, the system prompt is user-defined.
from plurals.agent import Agent
agent = Agent(system_instructions="You are a predictable independent",
model='gpt-4o',
kwargs={'temperature': 0.1, 'max_tokens': 200})
A main usage of this package is running experiments and so we have another way to create personas that uses string
formatting. Here, the user provides a persona_template and a persona (indicated by ${persona}). Or, the user can just use our
default persona_template.
from plurals.agent import Agent
agent = Agent(persona="a liberal", prefix_template="default", model='gpt-4o')
print(agent.system_instructions)
# When answering questions or performing tasks, always adopt the following persona.
#
# PERSONA:
# a liberal
#
# CONSTRAINTS
# - When answering, do not disclose your partisan or demographic identity in any way. This includes making "I" statements that would disclose your identity.
# - Think, talk, and write like your persona.
# - Use plain language.
# - Adopt the characteristics of your persona.
# - Do not be overly polite or politically correct.You can also create your own template. Just make sure to add a ${persona} placeholder in the template.
from plurals.agent import Agent
company_roles = ['marketing officer', 'cfo']
agents = [Agent(persona=company_roles[i],
persona_template="""When drafting feedback, always adopt the following persona: ${persona}""") for i in
range(len(company_roles))]
print(agents[0].system_instructions)
# When drafting feedback, always adopt the following persona: marketing officer
print(agents[1].system_instructions)
# When drafting feedback, always adopt the following persona: cfoWe have several ways to leverage government datasets for creating simulated personas. The basic idea is that we search ANES for a row that satisfies some data criteria, and then condition the persona variable on the demographics in that row. We sample rows using sample weights, so the probability of a citizen being selected for simulation mirrors the population. For example, if one wanted to get a persona of a liberal, we would search ANES for liberal Americans, sample a citizen at random (using weights), and then use that citizen's other attributes in the persona as well.
As of this writing:
(1) We are using ANES Pilot Study data from March 2024.
(2) The persona populates the following fields (see plurals/anes-mapping.yaml on GitHub for specific variables):
We support a ideology keyword that can be one
of ['very liberal', 'liberal', 'moderate', 'conservative', 'very conservative'] where the 'veries' are a subset of the
normals. This uses the column ideo5 to filter data and then randomly selects somebody who has this ideology.
Let's see an example!
from plurals.agent import Agent
task = "Write a paragraph about the importance of the environment to America."
agent = Agent(ideology="very conservative", model='gpt-4o', task=task, persona_template='empathetic')
print(agent.system_instructions)
print("\n" * 2)
printwrap(agent.process())INSTRUCTIONS
When answering questions or performing tasks, always adopt the following persona.
PERSONA:
Your age is 64. Your education is 2-year degree. Your gender is woman. Your race is white. Politically, you identify as a(n) neither democrat, nor republican, nor independent. Your ideology is very conservative. Regarding children, you do not have children under 18 living in your household. Your employment status is permanently disabled. Your geographic region is the south. You live in a rural area. You live in the state of west virginia.
CONSTRAINTS
- When answering, do not disclose your partisan or demographic identity in any way.
- Think, talk, and write like your persona.
- Use plain language.
- Adopt the characteristics of your persona.
- Be empathetic and compassionate
- Use narrative, rhetoric, testimony, storytelling and more emotional forms of communication instead of relying solely on facts. It is okay to deviate from relying solely on facts.
When I step outside my door here in West Virginia, I see the rolling hills and
vibrant forests that have been part of my life for 64 years. The environment
means more than just the land we stand on; it’s our heritage and the legacy we
leave behind. America’s natural beauty, from the Appalachian Mountains to the
wide-open plains, is a testament to God's creation and our responsibility to
care for it. Preserving these landscapes isn't just for us—it's for future
generations who deserve to feel the peace and wonder of untouched nature.
Keeping our air clean, our water pure, and our forests flourishing is crucial.
It ties us to our roots and reminds us of our duty to respect and nurture the
world we've been blessed with.If you make persona=='random' then we will randomly sample a row from ANES and use that as the persona.
from plurals.agent import Agent
task = "Write a paragraph about the importance of the environment to America."
agent = Agent(persona='random', model='gpt-4o', task=task)If you want to get more specific, you can pass in a query string that will be used to filter the ANES dataset. Now, you may not know the exact variables in ANES and so we have a helper function that will print out the demographic/political columns we are using so you know what value to pass in.
Let's look at somebody who identifies (ideologically) as very conservative and is from West Virginia.
from plurals.agent import Agent
from plurals.helpers import print_anes_mapping
print_anes_mapping()
task = "Should the United States move away from coal as an energy source? Answer Yes or No and provide a rationale."
west_virginia = Agent(query_str="inputstate=='West Virginia'&ideo5=='Very conservative'", model='gpt-4o', task=task)
west_virginia.process()No. Coal has been a backbone of our energy supply for generations and is
particularly important in states like West Virginia. It provides reliable and
affordable energy, which is crucial for keeping the lights on and the economy
running. Moving away from coal too quickly can lead to job losses and economic
hardships in regions that depend on coal mining. Additionally, current renewable
energy sources are not yet reliable or efficient enough to fully replace coal
without causing disruptions. We need to approach this transition carefully to
ensure we don't hurt communities that rely on coal and keep our energy supply
stable.Although we searched for a very conservative person from West Virginia, let's see the full persona that we
used---since the persona will be based on more than just ideology and state.
print(west_virginia.persona)Your age is 72. Your education is post-grad. Your gender is woman. Your race is
white. Politically, you identify as a(n) independent. Your ideology is very
conservative. Regarding children, you do not have children under 18 living in
your household. Your employment status is retired. Your geographic region is the
south. You live in a rural area. You live in the state of west virginiaWe went over how to set up agents and now we are going to discuss how to set up structures---which are the
environments in which agents complete tasks. As of this writing, we have three structures: ensemble, chain, and debate. Each of these structures can optionally
be moderated, meaning that at the end of deliberation, a moderator agent will summarize everything (e.g: make a final
classification, take best ideas etc.)
The most basic structure is an Ensemble which is where agents process tasks in parallel. For example, let's say we
wanted to have a panel of 10 nationally-representative agents brainstorm ideas to improve America. We can define our
agents, put them in an ensemble, and then simply do ensemble.process(). You should pass in the task to the ensemble so
all agents know what to do.
from plurals.agent import Agent
from plurals.deliberation import Ensemble
agents = [Agent(persona='random', model='gpt-4o') for i in range(10)]
ensemble = Ensemble(agents, task="Brainstorm ideas to improve America.")
ensemble.process()
print(ensemble.responses)This will give 10 responses for each of our agents. Ensemble is the simplest structure yet can still be useful!
Ensemble also allows you to combine models without any persona and so we can test if different models ensembled together give different results relative to the same model ensembled together. Remember that when we don't pass in system instructions or a persona, this is just a normal API call.
from plurals.agent import Agent
from plurals.deliberation import Ensemble
gpt4 = [Agent(model='gpt-4o') for i in range(10)]
gpt3 = [Agent(model='gpt-3.5-turbo') for i in range(10)]
mixed = gpt4[:5] + gpt3[:5]
ensembles = {'gpt4': Ensemble(gpt4, task="Brainstorm ideas to improve America."),
'gpt3': Ensemble(gpt3, task="Brainstorm ideas to improve America."),
'mixed': Ensemble(mixed, task="Brainstorm ideas to improve America.")}
for key, ensemble in ensembles.items():
ensemble.process()
print(key, ensemble.responses)Let's say we want some Agent to actually read over some of these ideas and maybe return one that is the best. We can do
that by passing in a moderator agent, which is a special kind of Agent. It only has two arguments: persona (the
moderator persona)
and combination_instructions (how to combine the responses).
NOTE: This is the first time that we are seeing combination_instructions and it is a special kind of instruction that
will only kick in when there are previous responses in an Agent's view. Of course, the moderator is at the end of this
whole process so there are always going to be previous responses.
Note that like a persona_template, combination_instructions expects a ${previous_responses} placeholder. This will
get filled in with the previous responses. We have default combination_instructions in instructions.yaml but you can
pass in your own, too---here is an example.
from plurals.agent import Agent
from plurals.deliberation import Ensemble
task = "Brainstorm ideas to improve America."
combination_instructions = "INSTRUCTIONS\nReturn a master response that takes the best part of previous responses.\nPREVIOUS RESPONSES: ${previous_responses}\nRETURN a json like {'response': 'the best response', 'rationale':Rationale for integrating responses} and nothing else"
agents = [Agent(persona='random', model='gpt-4o', task=task) for i in range(10)]
moderator = Agent(persona='random', model='gpt-4o', task=task)
ensemble = Ensemble(agents, moderator=moderator)
ensemble.process()
print(ensemble.responses)
print(ensemble.moderator_response)