Spaces:
Paused
Paused
File size: 30,502 Bytes
6a42990 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 | import json
from chevron import render
from tinytroupe.agent import TinyPerson
from tinytroupe.environment import TinyWorld
from tinytroupe.utils import LLMChat, indent_at_current_level
from tinytroupe.experimentation import logger
from tinytroupe import default
class Proposition:
MIN_SCORE = 0
MAX_SCORE = 9
def __init__(self, claim:str, target=None, include_personas:bool=False, first_n:int=None, last_n:int=None,
double_check:bool=False, use_reasoning_model:bool=False, precondition_function=None):
"""
Define a proposition as a (textual) claim about a target, which can be a TinyWorld, a TinyPerson or several of any.
The proposition's truth value can then either be checked as a boolean or computed as an integer score denoting the degree of truth.
Sometimes a proposition is better used in an implicative way, i.e., as a claim that is true or false depending on the context. For example, when
considering the latest agent action, the proposition might be applicable only to certain agent action types. To allow this,
this class allows to define a precondition function, which effectivelly turns a proposition `P` into `Precondition --> P`. This is logically equivalent to
`not P or Precondition`. In other words:
- if the precondition is true, then the proposition is evaluated normally (as a boolean or a score).
- if the precondition is false, then the proposition is always true (or with highest score).
- if the precondition is None, then the proposition is evaluated normally (as a boolean or a score).
Args:
claim (str): the claim of the proposition
target (TinyWorld, TinyPerson, list): the target or targets of the proposition. If not given, it will have to be specified later.
include_personas (bool): whether to include the persona specifications of the agents in the context
first_n (int): the number of first interactions to consider in the context
last_n (int): the number of last interactions (most recent) to consider in the context
double_check (bool): whether to ask the LLM to double check its answer. This tends to give more strict answers, but is slower and more expensive.
use_reasoning_model (bool): whether to use a reasoning model to evaluate the proposition
precondition_function (function): a Boolean function that indicates whether the proposition can be evaluated or not. This is useful to avoid evaluating propositions that are not relevant for the current context. If the precondition fails, the proposition is always interpreted as true (or with highest score). MUST have named arguments `target`, `additional_context`, and `claim_variables` (note: you can use a lambda for this too, e.g., `lambda target, additional_context, claim_variables: ...`).
"""
self.claim = claim
self.targets = self._target_as_list(target)
self.include_personas = include_personas
self.first_n = first_n
self.last_n = last_n
self.double_check = double_check
self.use_reasoning_model = use_reasoning_model
self.precondition_function = precondition_function
# the chat with the LLM is preserved until the proposition is re-evaluated. While it is available,
# the chat can be used to follow up on the proposition, e.g., to ask for more details about the evaluation.
self.llm_chat = None
self.value = None
self.justification = None
self.confidence = None
self.recommendations = None
def __copy__(self):
"""
Create a shallow copy of the proposition without any evaluation state.
Returns:
Proposition: A new proposition with the same configuration parameters.
"""
new_prop = Proposition(
claim=self.claim,
target=self.targets,
include_personas=self.include_personas,
first_n=self.first_n,
last_n=self.last_n,
double_check=self.double_check,
use_reasoning_model=self.use_reasoning_model,
precondition_function=self.precondition_function
)
return new_prop
def copy(self):
"""
Create a shallow copy of the proposition without any evaluation state.
Returns:
Proposition: A new proposition with the same configuration parameters.
"""
return self.__copy__()
def __call__(self, target=None, additional_context=None, claim_variables:dict={}, return_full_response:bool=False) -> bool:
return self.check(target=target, additional_context=additional_context, claim_variables=claim_variables, return_full_response=return_full_response)
def _check_precondition(self, target, additional_context:str, claim_variables:dict) -> bool:
"""
Check whether the proposition can be evaluated or not.
"""
if self.precondition_function is None:
return True
else:
return self.precondition_function(target=target, additional_context=additional_context, claim_variables=claim_variables)
def check(self, target=None, additional_context="No additional context available.", claim_variables:dict={}, return_full_response:bool=False) -> bool:
"""
Check whether the proposition holds for the given target(s).
"""
current_targets = self._determine_target(target)
if self._check_precondition(target=current_targets, additional_context=additional_context, claim_variables=claim_variables) == False:
self.value = True
self.justification = "The proposition is trivially true due to the precondition being false."
self.confidence = 1.0
self.full_evaluation_response = {"value": True, "justification": self.justification, "confidence": self.confidence}
else: # precondition is true or None
context = self._build_context(current_targets)
# might use a reasoning model, which could allow careful evaluation of the proposition.
model = self._model(self.use_reasoning_model)
#render self.claim using the claim_variables via chevron
rendered_claim = render(self.claim, claim_variables)
self.llm_chat = LLMChat(system_prompt="""
You are a system that evaluates whether a proposition is true or false with respect to a given context. This context
always refers to a multi-agent simulation. The proposition is a claim about the behavior of the agents or the state of their environment
in the simulation.
The context you receive can contain one or more of the following:
- the trajectory of a simulation of one or more agents. This means what agents said, did, thought, or perceived at different times.
- the state of the environment at a given time.
Your output **must**:
- necessarily start with the word "True" or "False";
- optionally be followed by a justification. Please provide a very detailed justifications, including very concrete and specific mentions to elements that contributed to reducing or increasing the score. Examples:
* WRONG JUSTIFICATION (too abstract) example: " ... the agent behavior did not comply with key parts of its specification, thus a reduced score ... "
* CORRECT JUSTIFICATION (very precise) example: " ... the agent behavior deviated from key parts of its specification, specifically: S_1 was not met because <reason>, ..., S_n was not met becasue <reason>. Thus, a reduced score ..."
For example, the output could be of the form: "True, because <HIGHLY DETAILED, CONCRETE AND SPECIFIC REASONS HERE>." or merely "True" if no justification is needed.
""",
user_prompt=f"""
Evaluate the following proposition with respect to the context provided. Is it True or False?
# Proposition
This is the proposition you must evaluate:
```
{indent_at_current_level(rendered_claim)}
```
# Context
The context you must consider is the following.
{indent_at_current_level(context)}
# Additional Context (if any)
{indent_at_current_level(additional_context)}
""",
output_type=bool,
enable_reasoning_step=True,
temperature=0.5,
frequency_penalty=0.0,
presence_penalty=0.0,
model=model)
self.value = self.llm_chat()
if self.double_check:
self.llm_chat.add_user_message("Are you sure? Please revise your evaluation to make is correct as possible.")
revised_value = self.llm_chat()
if revised_value != self.value:
logger.warning(f"The LLM revised its evaluation: from {self.value} to {revised_value}.")
self.value = revised_value
self.reasoning = self.llm_chat.response_reasoning
self.justification = self.llm_chat.response_justification
self.confidence = self.llm_chat.response_confidence
self.full_evaluation_response = self.llm_chat.response_json
# return the final result, either only the value or the full response
if not return_full_response:
return self.value
else:
return self.full_evaluation_response
def score(self, target=None, additional_context="No additional context available.", claim_variables:dict={}, return_full_response:bool=False) -> int:
"""
Compute the score for the proposition with respect to the given context.
"""
current_targets = self._determine_target(target)
if self._check_precondition(target=current_targets, additional_context=additional_context, claim_variables=claim_variables) == False:
self.value = self.MAX_SCORE
self.justification = "The proposition is trivially true due to the precondition being false."
self.confidence = 1.0
self.full_evaluation_response = {"value": self.value, "justification": self.justification, "confidence": self.confidence}
else: # precondition is true or None
# build the context with the appropriate targets
context = self._build_context(current_targets)
# might use a reasoning model, which could allow careful evaluation of the proposition.
model = self._model(self.use_reasoning_model)
#render self.claim using the claim_variables via chevron
rendered_claim = render(self.claim, claim_variables)
self.llm_chat = LLMChat(system_prompt=f"""
You are a system that computes an integer score (between {Proposition.MIN_SCORE} and {Proposition.MAX_SCORE}, inclusive) about how much a proposition is true or false with respect to a given context.
This context always refers to a multi-agent simulation. The proposition is a claim about the behavior of the agents or the state of their environment in the simulation.
The minimum score of {Proposition.MIN_SCORE} means that the proposition is completely false in all of the simulation trajectories, while the maximum score of {Proposition.MAX_SCORE} means that the proposition is completely true in all of the simulation trajectories. Intermediate scores are used to express varying degrees of partially met expectations. When assigning a score, follow these guidelines:
- If the data required to judge the proposition is not present, assign a score of {Proposition.MAX_SCORE}. That is to say, unless there is evidence to the contrary, the proposition is assumed to be true.
- The maximum score of {Proposition.MAX_SCORE} should be assigned when the evidence is as good as it can be. That is to say, all parts of the observed simulation trajectory support the proposition, no exceptions.
- The minimum score of {Proposition.MIN_SCORE} should be assigned when the evidence is as bad as it can be. That is to say, all parts of the observed simulation trajectory contradict the proposition, no exceptions.
- Intermediate scores should be assigned when the evidence is mixed. The intermediary score should be proportional to the balance of evidence, according to these bands:
0 = The proposition is without any doubt completely false;
1, 2, 3 = The proposition has little support and is mostly false;
4, 5 = The evidence is mixed, and the proposition is as much true as it is false;
6, 7, 8 = The proposition is well-supported and is mostly true;
9 = The proposition is without any doubt completely true.
- You should be very rigorous in your evaluation and, when in doubt, assign a lower score.
- If there are critical flaws in the evidence, you should move your score to a lower band entirely.
- If the provided context has inconsistent information, you **must** consider **only** the information that gives the lowest score, since we want to be rigorous and if necessary err to the lower end.
* If you are considering the relationship between an agent specification and a simulation trajectory, you should consider the worst possible interpretation of: the agent specification; the simulation trajectory; or the relationship between the two.
* These contradictions can appear anywhere in the context. When they do, you **always** adopt the worst possible inteprpretation, because we want to be rigorous and if necessary err to the lower end. It does not matter if the contradiction shows only very rarely, or if it is very small. It is still a contradiction and should be considered as such.
* DO NOT dismiss contradictions as specification errors. They are part of the evidence and should be considered as such. They **must** be **always** taken into account when computing the score. **Never** ignore them.
Additionally, whenever you are considering the relationship between an agent specification and a simulation trajectory, the following additional scoring guidelines apply:
- All observed behavior **must** be easily mapped back to clear elements of the agent specification. If you cannot do this, you should assign a lower score.
- Evaluate **each** relevant elements in the simulation trajectory (e.g., actions, stimuli) one by one, and assign a score to each of them. The final score is the average of all the scores assigned to each element.
The proposition you receive can contain one or more of the following:
- A statement of fact, which you will score.
- Additional context, which you will use to evaluate the proposition. In particular, it might refer or specify potentail parts
of similation trajectories for consideration. These might be formatted differently than what is given in the main context, so
make sure you read them carefully.
- Additional instructions on how to evaluate the proposition.
The context you receive can contain one or more of the following:
- the persona specifications of the agents in the simulation. That is to say, what the agents **are**, not what they are **doing**.
- the simulation trajectories of one or more agents. This means what agents said, did, thought, or perceived at different times.
These trajectories **are not** part of the persona specification.
- the state of the environment at a given time.
- additional context that can vary from simulation to simulation.
To interpret the simulation trajectories, use the following guidelines:
- Agents can receive stimuli and produce actions. You might be concerned with both or only one of them, depending on the specific proposition.
- Actions are clearly marked with the text "acts", e.g., "Agent A acts: [ACTION]". If it is not thus marked, it is not an action.
- Stimuli are denoted by "--> Agent name: [STIMULUS]".
Your output **must**:
- necessarily start with an integer between {Proposition.MIN_SCORE} and {Proposition.MAX_SCORE}, inclusive;
- be followed by a justification. Please provide a very detailed justifications, including very concrete and specific mentions to elements that contributed to reducing or increasing the score. Examples:
* WRONG JUSTIFICATION (too abstract) example: " ... the agent behavior did not comply with key parts of its specification, thus a reduced score ... "
* CORRECT JUSTIFICATION (very precise) example: " ... the agent behavior deviated from key parts of its specification, specifically: S_1 was not met because <reason>, ..., S_n was not met becasue <reason>. Thus, a reduced score ..."
For example, the output could be of the form: "1, because <HIGHLY DETAILED, CONCRETE AND SPECIFIC REASONS HERE>."
""",
user_prompt=f"""
Compute the score for the following proposition with respect to the context provided. Think step-by-step to assign the most accurate score and provide a justification.
# Proposition
This is the proposition you must evaluate:
```
{indent_at_current_level(rendered_claim)}
```
# Context
The context you must consider is the following.
{indent_at_current_level(context)}
# Additional Context (if any)
{indent_at_current_level(additional_context)}
""",
output_type=int,
enable_reasoning_step=True,
temperature=1.0,
frequency_penalty=0.0,
presence_penalty=0.0,
# Use a reasoning model, which allows careful evaluation of the proposition.
model=model)
self.value = self.llm_chat()
if self.double_check:
self.llm_chat.add_user_message("Are you sure? Please revise your evaluation to make is correct as possible.")
revised_value = self.llm_chat()
if revised_value != self.value:
logger.warning(f"The LLM revised its evaluation: from {self.value} to {revised_value}.")
self.value = revised_value
self.reasoning = self.llm_chat.response_reasoning
self.justification = self.llm_chat.response_justification
self.confidence = self.llm_chat.response_confidence
self.full_evaluation_response = self.llm_chat.response_json
# return the final result, either only the value or the full response
if not return_full_response:
return self.value
else:
return self.full_evaluation_response
def recommendations_for_improvement(self):
"""
Get recommendations for improving the proposition.
"""
# TODO this is not working, let's try something else
#
#if self.llm_chat is None:
# raise ValueError("No evaluation has been performed yet. Please evaluate the proposition before getting recommendations.")
#
#self.llm_chat.add_system_message(\
# """
# You will now act as a system that provides recommendations for the improvement of the scores previously assigned to propositions.
# You will now output text that contains analysises, recommendations and other information as requested by the user.
# """)
#
#self.llm_chat.add_user_message(\
# """
# To help improve the score next time, please list the following in as much detail as possible:
# - all recommendations for improvements based on the current score.
# - all criteria you are using to assign scores, and how to best satisfy them
#
# For both cases:
# - besides guidelines, make sure to provide plenty of concrete examples of what to be done in order to maximize each criterion.
# - avoid being generic or abstract. Instead, all of your criteria and recommendations should be given in very concrete terms that would work specifically for the case just considered.
#
# Note that your output is a TEXT with the various recommendations, information and tips, not a JSON object.
#
# Recommendations:
# """)
#
#recommendation = self.llm_chat(output_type=str, enable_json_output_format=False)
recommendation = "No additional recommendations at this time."
return recommendation
def _model(self, use_reasoning_model):
if use_reasoning_model:
return default["reasoning_model"]
else:
return default["model"]
def _determine_target(self, target):
"""
Determine the target for the proposition. If a target was provided during initialization, it must not be provided now (i.e., the proposition is immutable).
If no target was provided during initialization, it must be provided now.
"""
# If no target was provided during initialization, it must be provided now.
if self.targets is None :
if target is None:
raise ValueError("No target specified. Please provide a target.")
else:
return self._target_as_list(target)
# If it was provided during initialization, it must not be provided now (i.e., the proposition is immutable).
else:
if target is not None:
raise ValueError("Target already specified. Please do not provide a target.")
else:
return self.targets
def _build_context(self, current_targets):
#
# build the context with the appropriate targets
#
context = ""
for target in current_targets:
target_trajectory = target.pretty_current_interactions(max_content_length=None, first_n=self.first_n, last_n=self.last_n)
if isinstance(target, TinyPerson):
if self.include_personas:
context += f"## Agent '{target.name}' Persona Specification\n\n"
context += "Before presenting the actual simulation trajectory, here is the persona specification of the agent that was used to produce the simulation.\n\n"
context += "This IS NOT the actual simulation, but only the static persona specification of the agent.\n\n"
context += f"persona={json.dumps(target._persona, indent=4)}\n\n"
context += f"## Agent '{target.name}' Simulation Trajectory (if any)\n\n"
elif isinstance(target, TinyWorld):
if self.include_personas:
context += f"## Environment '{target.name}' Personas Specifications\n\n"
context += "Before presenting the actual simulation trajectory, here are the persona specifications of the agents used to produce the simulation.\n\n"
context += "This IS NOT the actual simulation, but only the static persona specification of the agent.\n\n"
for agent in target.agents:
context += f"### Agent '{agent.name}' Persona Specification\n\n"
context += f"persona={json.dumps(agent._persona, indent=4)}\n\n"
context += f"## Environment '{target.name}' Simulation Trajectory (if any)\n\n"
context += target_trajectory + "\n\n"
return context
def _target_as_list(self, target):
if target is None:
return None
elif isinstance(target, TinyWorld) or isinstance(target, TinyPerson):
return [target]
elif isinstance(target, list) and all(isinstance(t, TinyWorld) or isinstance(t, TinyPerson) for t in target):
return target
else:
raise ValueError("Target must be a TinyWorld, a TinyPerson or a list of them.")
def check_proposition(target, claim:str, additional_context="No additional context available.",
first_n:int=None, last_n:int=None,
return_full_response:bool=False):
"""
Check whether a propositional claim holds for the given target(s). This is meant as a
convenience method to avoid creating a Proposition object (which you might not need
if you are not interested in the justification or confidence of the claim, or will
not use it again).
Args:
target (TinyWorld, TinyPerson, list): the target or targets of the proposition
claim (str): the claim of the proposition
additional_context (str): additional context to provide to the LLM
first_n (int): the number of first interactions to consider in the context
last_n (int): the number of last interactions (most recent) to consider in the context
return_full_response (bool): whether to return the full response from the LLM, including justification and confidence
Returns:
bool: whether the proposition holds for the given target(s)
"""
proposition = Proposition(claim, target, first_n=first_n, last_n=last_n)
return proposition.check(additional_context=additional_context, return_full_response=return_full_response)
def compute_score(target, claim:str,
additional_context="No additional context available.",
first_n:int=None, last_n:int=None,
return_full_response:bool=False):
"""
Compute a score about whether a claim holds for the given target(s). This is meant as a
convenience method to avoid creating a Score object (which you might not need
if you are not interested in the justification or confidence of the claim, or will
not use it again).
Args:
target (TinyWorld, TinyPerson, list): the target or targets of the proposition
claim (str): the claim of the proposition
additional_context (str): additional context to provide to the LLM
first_n (int): the number of first interactions to consider in the context
last_n (int): the number of last interactions (most recent) to consider in the context
return_full_response (bool): whether to return the full response from the LLM, including justification and confidence
Returns:
bool: whether the proposition holds for the given target(s)
"""
score = Proposition(claim, target,
first_n=first_n, last_n=last_n)
return score.compute(additional_context=additional_context, return_full_response=return_full_response) |