From e751763b46b133c407097471b0ab636e79778fbd Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Daniel=20Grie=C3=9Fhaber?= <griesshaber@hdm-stuttgart.de>
Date: Tue, 20 Aug 2024 08:42:28 +0200
Subject: [PATCH] add interactive editor to correct badly judged reponses
---
evoprompt/api/backend.py | 1 -
evoprompt/evolution.py | 51 +++++++++++++-----
evoprompt/models.py | 41 +++++----------
evoprompt/optimization.py | 108 ++++++++++++++++++++------------------
evoprompt/utils.py | 2 +-
5 files changed, 109 insertions(+), 94 deletions(-)
diff --git a/evoprompt/api/backend.py b/evoprompt/api/backend.py
index 608d26b..2c9e0eb 100644
--- a/evoprompt/api/backend.py
+++ b/evoprompt/api/backend.py
@@ -74,7 +74,6 @@ class MultiProcessOptimizer:
self.model_exec = None
def _setup_models(self) -> None:
- print("setup models")
if self._evolution_model is not None:
raise Exception("Evolution model has already been initialized.")
diff --git a/evoprompt/evolution.py b/evoprompt/evolution.py
index ec67579..2b4e6c7 100644
--- a/evoprompt/evolution.py
+++ b/evoprompt/evolution.py
@@ -8,7 +8,7 @@ from tqdm import trange
from evoprompt.cli import argument_parser
from evoprompt.models import LLMModel
from evoprompt.opt_types import ModelUsage, Prompt
-from evoprompt.optimization import PromptOptimization
+from evoprompt.optimization import Judgement, PromptOptimization
from evoprompt.task import Task
from evoprompt.template_de import get_de_prompt_template
from evoprompt.utils import get_all_subclasses, get_rng, log_calls
@@ -85,7 +85,7 @@ class EvolutionAlgorithm(PromptOptimization, metaclass=ABCMeta):
*,
prompts_current_evolution: list[Prompt],
current_iteration: int,
- ):
+ ) -> tuple[str, list[Judgement], ModelUsage]:
pass
@abstractmethod
@@ -116,7 +116,11 @@ class EvolutionAlgorithm(PromptOptimization, metaclass=ABCMeta):
# Line 4: Evolution: generate a new prompt based on the selected parent prompts by leveraging LLM to perform evolutionary operators
# p′i â†Evo(pr1,...,prk)
- p_i, evolution_usage = self.evolve(
+ (
+ p_i,
+ judgements,
+ evolution_usage,
+ ) = self.evolve(
pr1,
pr2,
prompts_current_evolution=prompts_current_evolution,
@@ -124,7 +128,14 @@ class EvolutionAlgorithm(PromptOptimization, metaclass=ABCMeta):
)
self.total_evolution_usage += evolution_usage
- evolved_prompt = self.add_prompt(p_i, (pr1, pr2), {"gen": t})
+ prompt_source = (
+ "corrected" if not all(j.happy for j in judgements) else "generated"
+ )
+ evolved_prompt = self.add_prompt(
+ p_i,
+ parents=(pr1, pr2),
+ meta={"gen": t, "source": prompt_source, "judgements": judgements},
+ )
self.total_evaluation_usage += evolved_prompt.usage
new_evolutions.append(evolved_prompt)
@@ -171,7 +182,7 @@ class GeneticAlgorithm(EvolutionAlgorithm):
prompt_1: str,
prompt_2: str,
**kwargs,
- ) -> tuple[str, ModelUsage]:
+ ):
# Following the evolutionary operators in GA, a new candidate prompt is generated through
# a two-step process based on the selected two parents:
# 1) The parent prompts undergo crossover, resulting in a new prompt that
@@ -186,6 +197,10 @@ class GeneticAlgorithm(EvolutionAlgorithm):
system_message=SYSTEM_MESSAGE,
prompt=filled_prompt,
)
+
+ judgement = self.judge_and_correct_step(filled_prompt, evolved_prompt, messages)
+ evolved_prompt = judgement.corrected_response
+
if "<prompt>" in evolved_prompt:
evolved_prompt = evolved_prompt.split("<prompt>")[1].split("</prompt>")[0]
@@ -195,9 +210,8 @@ class GeneticAlgorithm(EvolutionAlgorithm):
prompt_2,
evolved_prompt,
)
- self.judge_step(filled_prompt, evolved_prompt, messages)
- return evolved_prompt, usage
+ return evolved_prompt, [judgement], usage
@log_calls("Performing update for GA")
def update(
@@ -255,6 +269,10 @@ class DifferentialEvolution(EvolutionAlgorithm):
system_message=SYSTEM_MESSAGE,
prompt=filled_prompt,
)
+
+ judgement = self.judge_and_correct_step(filled_prompt, evolved_prompt, messages)
+ evolved_prompt = judgement.corrected_response
+
matches = re.findall(
# regex that matches any characters between last pair of <prompt></prompt>, also if </prompt> is missing
r"<prompt>(?!.*<prompt>)(?:(.*)</prompt>|(.*))",
@@ -281,9 +299,7 @@ class DifferentialEvolution(EvolutionAlgorithm):
evolved_prompt,
)
- self.judge_step(filled_prompt, evolved_prompt, messages)
-
- return evolved_prompt, usage
+ return evolved_prompt, [judgement], usage
@log_calls("Performing update for DE")
def update(
@@ -322,6 +338,9 @@ class DifferentialEvolutionWithCot(DifferentialEvolution):
)
messages = None
+ response: str = ""
+ judgements: list[Judgement] = []
+ usage: ModelUsage = ModelUsage()
for idx, prompt in enumerate(DE_COT_PROMPTS):
filled_prompt = prompt.format(
prompt1=prompt_1,
@@ -341,9 +360,13 @@ class DifferentialEvolutionWithCot(DifferentialEvolution):
messages,
response,
)
- self.judge_step(filled_prompt, response, history=messages)
- # input(messages)
- # input(response)
+ judgement = self.judge_and_correct_step(
+ filled_prompt, response, history=messages
+ )
+ judgements.append(judgement)
+ # replace last message with corrected response
+ messages[-1]["content"] = judgement.corrected_response
+ response = judgement.corrected_response
# at this point we should get a new prompt
if "<prompt>" in response:
@@ -358,7 +381,7 @@ class DifferentialEvolutionWithCot(DifferentialEvolution):
response,
)
- return response, usage
+ return response, judgements, usage
def get_all_subclasses(cls):
diff --git a/evoprompt/models.py b/evoprompt/models.py
index b2877e7..a10005d 100644
--- a/evoprompt/models.py
+++ b/evoprompt/models.py
@@ -63,24 +63,11 @@ class LLMModel(ABC):
prompt: str,
*,
use_cache: bool = False,
- stop: str = None,
+ stop: str | None = None,
history: ChatMessages | None = None,
**kwargs: Any,
- ) -> tuple[str, list[dict[str, str]], ModelUsage]:
- # create prompt
- prompt = prompt_prefix + prompt + prompt_suffix + prompt_appendix
- messages = [self._get_user_message(prompt)]
- model_input = self.build_model_input(prompt, system_message, messages, history)
-
- reponse, usage = self._create_completion(
- **model_input,
- stop=stop,
- use_cache=use_cache,
- **kwargs,
- )
-
- messages.append(self._get_assistant_message(reponse))
- return reponse, messages, usage
+ ) -> tuple[str, ChatMessages, ModelUsage]:
+ pass
def _get_user_message(self, content: str):
return {
@@ -193,7 +180,7 @@ class Llama(LLMModel):
prompt: str,
*,
use_cache: bool = False,
- stop: str = None,
+ stop: str | None = None,
history: ChatMessages | None = None,
**kwargs: Any,
) -> tuple[str, ModelUsage]:
@@ -274,22 +261,22 @@ class ChatModel:
prompt: str,
*,
use_cache: bool = False,
- stop: str = None,
+ stop: str | None = None,
history: ChatMessages | None = None,
**kwargs: Any,
) -> tuple[str, ModelUsage]:
# a history is prepended to the messages, and we assume that it also includes a system message, i.e., we never add a system message in this case
# TODO is it better to check for a system message in the history?
- if history is None and system_message:
- messages = [self._get_system_message(system_message)]
- elif history is not None:
- messages = history
- else:
- messages = []
- messages += [self._get_user_message(prompt)]
+ messages = [self._get_user_message(prompt)]
+
+ if history is None:
+ if system_message:
+ history = [self._get_system_message(system_message)]
+ else:
+ history = []
reponse, usage = self._create_completion(
- messages=messages,
+ messages=history + messages,
stop=stop,
use_cache=use_cache,
max_tokens=self.options.max_tokens,
@@ -297,7 +284,7 @@ class ChatModel:
)
messages.append(self._get_assistant_message(reponse))
- return reponse, messages, usage
+ return reponse, history + messages, usage
class LlamaChat(ChatModel, Llama):
diff --git a/evoprompt/optimization.py b/evoprompt/optimization.py
index 648325e..4f91aba 100644
--- a/evoprompt/optimization.py
+++ b/evoprompt/optimization.py
@@ -1,16 +1,14 @@
import json
import logging
-from abc import abstractmethod
-from itertools import zip_longest
from pathlib import Path
-from typing import Any, Optional
+from typing import Any, Literal, NamedTuple, Optional, TypedDict
from textual.app import App, ComposeResult
from textual.binding import Binding
-from textual.widgets import Footer, TextArea
+from textual.widgets import Collapsible, Footer, Label, TextArea
from tqdm import tqdm, trange
-from evoprompt.models import LLMModel
+from evoprompt.models import ChatMessages, LLMModel
from evoprompt.opt_types import ModelUsage, OptTypeEncoder, Prompt
from evoprompt.task import Task
from evoprompt.utils import initialize_run_directory, log_calls
@@ -20,19 +18,48 @@ logger = logging.getLogger(__name__)
PARAPHRASE_PROMPT = """You are given an instruction that describes a task. Write a response that paraphrases the instruction. Only output the paraphrased instruction bracketed in <prompt> and </prompt>."""
+PromptSource = Literal["baseprompt", "paraphrase", "evolution", "corrected"]
-class EditText(App):
+
+class Judgement(NamedTuple):
+ original_response: str
+ corrected_response: str
+ happy: bool
+
+
+class PromptMeta(TypedDict):
+ gen: int
+ source: PromptSource
+ judgements: list[Judgement]
+
+
+class ResponseEditor(App):
BINDINGS = [
- Binding(key="q", action="quit", description="Quit the app"),
+ Binding(
+ key="ctrl+q", action="quit", description="Finish Editing & Save Prompt"
+ ),
]
- def __init__(self, text):
- self.text = text + "asdasdasd"
+ def __init__(self, instruction: str, original_response: str, history: ChatMessages):
+ self.instruction = instruction
+ self.response = original_response
+ self.history = history
super().__init__()
def compose(self) -> ComposeResult:
- self.text_area = TextArea.code_editor(self.text)
- yield Footer(self.text_area)
+ self.text_area = TextArea.code_editor(self.response)
+ for idx, message in enumerate(self.history[:-1]):
+ yield Collapsible(
+ Label(message["content"]),
+ title=message["role"],
+ collapsed=idx != len(self.history) - 2,
+ )
+ yield self.text_area
+ yield Footer()
+
+ @property
+ def modified_response(self):
+ return self.text_area.text
@log_calls("Paraphrasing prompts")
@@ -133,6 +160,7 @@ class PromptOptimization:
prompt for _, prompt in sorted_results[: num_initial_prompts // 2]
]
initial_population = top_prompts.copy()
+ prompt_sources = ["baseprompt" for _ in initial_population]
# fill up the rest with paraphrases of the top prompts
promptindex_to_paraphrase = 0
@@ -146,6 +174,7 @@ class PromptOptimization:
)
self.total_evolution_usage += paraphrase_usage
initial_population += paraphrases
+ prompt_sources.append("paraphrase")
logger.info(
"Paraphrased prompt '%s': %s.",
top_prompts[promptindex_to_paraphrase]
@@ -156,13 +185,10 @@ class PromptOptimization:
promptindex_to_paraphrase += 1
promptindex_to_paraphrase %= len(top_prompts)
- return initial_population
+ return initial_population, prompt_sources
def add_prompt(
- self,
- prompt: str,
- parents: tuple[Prompt] | None = None,
- meta: dict | None = None,
+ self, prompt: str, parents: tuple[Prompt] | None, meta: PromptMeta
) -> Prompt:
score, usage, history = self.evaluate_prompt(prompt, parents)
prompt_object = Prompt(
@@ -188,26 +214,12 @@ class PromptOptimization:
return prompt_object
- def add_prompts(
- self,
- prompts: list[str],
- parents: list[tuple[Prompt]] = iter(()),
- metas: list[dict] = iter(()),
- ):
- return [
- self.add_prompt(prompt, _parents, meta)
- for prompt, _parents, meta in zip_longest(prompts, parents, metas)
- ]
-
def get_prompt(self, prompt_id: str):
return self.all_prompts[prompt_id]
def get_prompts(self, prompt_ids: list[str]):
return [self.get_prompt(p_id) for p_id in prompt_ids]
- @abstractmethod
- def save_snapshot(self): ...
-
def init_run(self, num_initial_prompts: int, num_iterations: int):
# family_tree contains the relation of prompts to its parents
self.family_tree: dict[str, tuple[str, ...] | None] = {}
@@ -224,16 +236,15 @@ class PromptOptimization:
self.save_snapshot()
# the initial prompts
- initial_prompts = self.get_initial_prompts(num_initial_prompts)
- initial_prompts = self.add_prompts(
- initial_prompts, metas=[{"gen": 0} for _ in initial_prompts]
- )
- # - Initial prompts P0 = {p1, p2, . . . , pN }
- self.P.append(initial_prompts)
-
- # accumulate usage
- for prompt in initial_prompts:
+ initial_prompts, prompt_sources = self.get_initial_prompts(num_initial_prompts)
+ self.P.append([])
+ for prompt, prompt_source in zip(initial_prompts, prompt_sources):
+ prompt = self.add_prompt(
+ prompt, parents=None, meta={"gen": 0, "source": prompt_source}
+ )
self.total_evaluation_usage += prompt.usage
+ # - Initial prompts P0 = {p1, p2, . . . , pN }
+ self.P[0].append(prompt)
self.save_snapshot()
def save_snapshot(self):
@@ -264,14 +275,9 @@ class PromptOptimization:
cls=OptTypeEncoder,
)
- def judge_step(
- self, instruction: str, response: str, history: list[dict[str, str]]
- ):
- print("\n" * 20)
- print(
- "Instruction:", instruction, "\nResponse:", response, "\nHistory:", history
- )
-
+ def judge_and_correct_step(
+ self, instruction: str, response: str, history: ChatMessages
+ ) -> Judgement:
# TODO: judge the actual response
judge_happy = False
@@ -279,14 +285,14 @@ class PromptOptimization:
f"{self.judge_model.__class__.__name__} judged the response as {'good' if judge_happy else 'bad'}"
)
if judge_happy:
- return True
+ return Judgement(response, response, happy=True)
logger.info(f"Prompt judged as bad. Letting User change the prompt.")
- editor = EditText(response)
+ editor = ResponseEditor(instruction, response, history)
editor.run()
- response = editor.text_area.text
- print(response)
+
+ return Judgement(response, editor.modified_response, happy=False)
def load_snapshot(path: Path):
diff --git a/evoprompt/utils.py b/evoprompt/utils.py
index 7cb3e1b..755b796 100644
--- a/evoprompt/utils.py
+++ b/evoprompt/utils.py
@@ -65,7 +65,7 @@ RUNS_DIR = current_directory.parent / "runs"
RUNS_DIR.mkdir(exist_ok=True)
-def initialize_run_directory(model: Callable):
+def initialize_run_directory(model):
global file_handler
# get package logger
--
GitLab