Moderators#

Moderators are a subclass of Agents who summarize multi-agent deliberation. Any Structure supports an optional Moderator. Moderators are defined by:

Profile: Like Agents, Moderators have a distinct “profile” which we operationalize as system instructions. System instructions can be set directly or via persona methods. We have a special class of Moderators called “Auto-Moderators” who generate their own system instructions based on a task.
Combination Instructions: Here, combination instructions define how Moderators aggregate the responses that they see. As with Agents, these can be set directly or via templates. These templates can be found in instructions.yaml.
Task: Moderators can have a distinct task from Agents, or inherit the task from the Structure they are moderating.
Model: Moderators are initialized to be a particular LLM.

Setting a Moderator’s system instructions#

Via personas#

Like Agents, personas and system_instructions are different ways to set up the moderator’s system instructions. If you use persona, then you can use some of our default moderator personas available in the defaults file. For example, if we pass in persona='voting', then we will use a moderator persona meant for voting.

from plurals.deliberation import Moderator

a = Moderator(persona='voting', model='gpt-4o', combination_instructions="voting")

There is also an option to define your own persona. However, when passing your own instructions in, note that, like persona_template, persona expects a ${task} placeholder. This will get filled in with the actual task.

from plurals.deliberation import Moderator

mod = Moderator(persona="You are a neutral moderator overseeing this task, ${task}", model='gpt-4o',
combination_instructions="voting")

Moderator system instructions set directly#

You can also set system_instructions directly, much like with Agents, and this will have a similar effect to custom personas.

from plurals.deliberation import Moderator

mod = Moderator(system_instructions="You are a neutral moderator overseeing this task, ${task}", model='gpt-4o',
combination_instructions="voting")

The difference is that system_instructions is not linked with our templates, so you cannot access any pre-defined moderator instructions when using system_instructions. system_instructions='default' will not access the default template like persona='default' would.

Auto-moderators#

We have a special option where, if the system_instructions of a moderator are set to auto, then the moderator will, given a task, come up with its own system instructions. Here is how to do this!

from plurals.deliberation import Moderator, Ensemble, Chain
from plurals.agent import Agent

task = ("Your goal is to come up with the most creative ideas possible for pants. We are maximizing creativity. Answer"
        " in 20 words.")
a = Agent(model='gpt-4o')
b = Agent(model='gpt-3.5-turbo')
# By putting the moderator in the Ensemble we are going to
# trigger the auto-mod generator
chain = Chain([a, b], moderator=Moderator(system_instructions='auto', model='gpt-4o'), task=task)

So let’s see what the moderator thinks it should be doing with this information.

print(chain.moderator.system_instructions)

Review all submitted responses, identify the most unique and imaginative ideas
for pants, and compile a ranked list from most to least creative. Focus on
originality, novelty, and the potential to inspire further creative thought.
Exclude repetitive or overly conventional ideas.

Here are different ways to initialize auto-moderation.

from plurals.deliberation import Moderator, Ensemble, Chain
from plurals.agent import Agent
task = "Come up with creative uses for a bat"

# Initializing the moderator with auto system instructions
a = Agent(model='gpt-4o')
b = Agent(model='gpt-3.5-turbo')
mod = Moderator(system_instructions='auto', model='gpt-4o', task=task)
chain = Ensemble([a, b], moderator=mod, task=task)


# Simply defining the moderator in the Structure will inherit the structure's task so this is also a simple way to have
# the Moderator bootstrap its own instructions based on the task.
a = Agent(model='gpt-4o')
b = Agent(model='gpt-3.5-turbo')
chain = Chain([a, b], moderator=Moderator(system_instructions='auto', model='gpt-4o'), task=task)


# You can also turn a normal moderator into an auto-moderator.
mod = Moderator(system_instructions="some boring initial instructions",  model='gpt-4o')
mod.generate_and_set_system_instructions(task=task)

# Or, you can generate instructions and inspect them before setting them. You can generate multiple times of course.
mod = Moderator(system_instructions="some boring initial instructions",  model='gpt-4o')
print(mod.generate_system_instructions(task=task))

# Review all submitted responses for uniqueness, relevance, and creativity. Prioritize ideas that are feasible and
# innovative. Eliminate duplicates and rank responses by feasibility and impact. Summarize the top 5-10 ideas,
# ensuring a diverse range of concepts is represented.

Now we set the instructions.

# Review all submitted responses for uniqueness, relevance, and creativity. Prioritize ideas that are feasible and
# innovative. Eliminate duplicates and rank responses by feasibility and impact. Summarize the top 5-10 ideas,
# ensuring a diverse range of concepts is represented.
mod.system_instructions = "Review all submitted responses for uniqueness, relevance, and creativity. Prioritize ideas that are feasible and innovative. Eliminate duplicates and rank responses by feasibility and impact. Summarize the top 5-10 ideas, ensuring a diverse range of concepts is represented."

Setting a Moderator’s combination instructions#

Combination instructions are set the same way as regular Agents, where you can use a template or input your own combination instructions that have a ${previous_responses} placeholder.

from plurals.deliberation import Moderator

mod = Moderator(system_instructions="You are a neutral moderator overseeing this task, ${task}", model='gpt-4o',
combination_instructions="Select the response from previous responses that is the least polarizing: <start>${previous_responses}</end>"))

Moderators

Contents