File size: 23,832 Bytes
f5776d3 |
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 |
import dsp
import dspy
from dspy.teleprompt.teleprompt import Teleprompter
from dspy.signatures import Signature
from dspy.evaluate.evaluate import Evaluate
from collections import defaultdict
import random
from dspy.teleprompt import BootstrapFewShot
import numpy as np
import optuna
import math
"""
USAGE SUGGESTIONS:
The following code can be used to compile a optimized signature teleprompter using the BayesianSignatureOptimizer, and evaluate it on an end task:
from dspy.teleprompt import BayesianSignatureOptimizer
teleprompter = BayesianSignatureOptimizer(prompt_model=prompt_model, task_model=task_model, metric=metric, n=10, init_temperature=1.0)
kwargs = dict(num_threads=NUM_THREADS, display_progress=True, display_table=0)
compiled_prompt_opt = teleprompter.compile(program, devset=devset[:DEV_NUM], optuna_trials_num=100, max_bootstrapped_demos=3, max_labeled_demos=5, eval_kwargs=kwargs)
eval_score = evaluate(compiled_prompt_opt, devset=evalset[:EVAL_NUM], **kwargs)
Note that this teleprompter takes in the following parameters:
* prompt_model: The model used for prompt generation. When unspecified, defaults to the model set in settings (ie. dspy.settings.configure(lm=task_model)).
* task_model: The model used for prompt generation. When unspecified, defaults to the model set in settings (ie. dspy.settings.configure(lm=task_model)).
* metric: The task metric used for optimization.
* n: The number of new prompts and sets of fewshot examples to generate and evaluate. Default=10.
* init_temperature: The temperature used to generate new prompts. Higher roughly equals more creative. Default=1.0.
* verbose: Tells the method whether or not to print intermediate steps.
* track_stats: Tells the method whether or not to track statistics about the optimization process.
If True, the method will track a dictionary with a key corresponding to the trial number,
and a value containing a dict with the following keys:
* program: the program being evaluated at a given trial
* score: the last average evaluated score for the program
* pruned: whether or not this program was pruned
This information will be returned as attributes of the best program.
"""
class BasicGenerateInstruction(Signature):
"""You are an instruction optimizer for large language models. I will give you a ``signature`` of fields (inputs and outputs) in English. Your task is to propose an instruction that will lead a good language model to perform the task well. Don't be afraid to be creative."""
basic_instruction = dspy.InputField(desc="The initial instructions before optimization")
proposed_instruction = dspy.OutputField(desc="The improved instructions for the language model")
proposed_prefix_for_output_field = dspy.OutputField(desc="The string at the end of the prompt, which will help the model start solving the task")
class BasicGenerateInstructionWithDataObservations(Signature):
"""You are an instruction optimizer for large language models. I will give you a ``signature`` of fields (inputs and outputs) in English. I will also give you some ``observations`` I have made about the dataset and task. Your task is to propose an instruction that will lead a good language model to perform the task well. Don't be afraid to be creative."""
basic_instruction = dspy.InputField(desc="The initial instructions before optimization")
observations = dspy.InputField(desc="Observations about the dataset and task")
proposed_instruction = dspy.OutputField(desc="The improved instructions for the language model")
proposed_prefix_for_output_field = dspy.OutputField(desc="The string at the end of the prompt, which will help the model start solving the task")
class BasicGenerateInstructionWithExamples(dspy.Signature):
("""You are an instruction optimizer for large language models. I will give you a ``signature`` of fields (inputs and outputs) in English. Specifically, I will also provide you with the current ``basic instruction`` that is being used for this task. I will also provide you with some ``examples`` of the expected inputs and outputs.
Your task is to propose an instruction that will lead a good language model to perform the task well. Don't be afraid to be creative.""")
# attempted_instructions = dspy.InputField(format=str, desc="Previously attempted task instructions, along with their resulting validation score, and an example of the instruction in use on a sample from our dataset.")
basic_instruction = dspy.InputField(desc="The initial instructions before optimization")
# examples = dspy.InputField(format=dsp.passages2text, desc="Example(s) of the task")
examples = dspy.InputField(format=dsp.passages2text, desc="Example(s) of the task")
proposed_instruction = dspy.OutputField(desc="The improved instructions for the language model")
proposed_prefix_for_output_field = dspy.OutputField(desc="The string at the end of the prompt, which will help the model start solving the task")
class BasicGenerateInstructionWithExamplesAndDataObservations(dspy.Signature):
("""You are an instruction optimizer for large language models. I will give you a ``signature`` of fields (inputs and outputs) in English. Specifically, I will also provide you with the current ``basic instruction`` that is being used for this task. I will also provide you with some ``observations`` I have made about the dataset and task, along with some ``examples`` of the expected inputs and outputs.
Your task is to propose a new improved instruction and prefix for the output field that will lead a good language model to perform the task well. Don't be afraid to be creative.""")
basic_instruction = dspy.InputField(desc="The initial instructions before optimization")
observations = dspy.InputField(desc="Observations about the dataset and task")
examples = dspy.InputField(format=dsp.passages2text, desc="Example(s) of the task")
proposed_instruction = dspy.OutputField(desc="The improved instructions for the language model")
proposed_prefix_for_output_field = dspy.OutputField(desc="The string at the end of the prompt, which will help the model start solving the task")
class ObservationSummarizer(dspy.Signature):
("""Given a series of observations I have made about my dataset, please summarize them into a brief 2-3 sentence summary which highlights only the most important details.""")
observations = dspy.InputField(desc="Observations I have made about my dataset")
summary = dspy.OutputField(desc="Two to Three sentence summary of only the most significant highlights of my observations")
class DatasetDescriptor(dspy.Signature):
("""Given several examples from a dataset please write observations about trends that hold for most or all of the samples. """
"""Some areas you may consider in your observations: topics, content, syntax, conciceness, etc. """
"""It will be useful to make an educated guess as to the nature of the task this dataset will enable. Don't be afraid to be creative""")
examples = dspy.InputField(desc="Sample data points from the dataset")
observations = dspy.OutputField(desc="Somethings that holds true for most or all of the data you observed")
class DatasetDescriptorWithPriorObservations(dspy.Signature):
("""Given several examples from a dataset please write observations about trends that hold for most or all of the samples. """
"""I will also provide you with a few observations I have already made. Please add your own observations or if you feel the observations are comprehensive say 'COMPLETE' """
"""Some areas you may consider in your observations: topics, content, syntax, conciceness, etc. """
"""It will be useful to make an educated guess as to the nature of the task this dataset will enable. Don't be afraid to be creative""")
examples = dspy.InputField(desc="Sample data points from the dataset")
prior_observations = dspy.InputField(desc="Some prior observations I made about the data")
observations = dspy.OutputField(desc="Somethings that holds true for most or all of the data you observed or COMPLETE if you have nothing to add")
class BayesianSignatureOptimizer(Teleprompter):
def __init__(self, prompt_model=None, task_model=None, teacher_settings={}, n=10, metric=None, init_temperature=1.0, verbose=False, track_stats=False, view_data_batch_size=10):
self.n = n
self.metric = metric
self.init_temperature = init_temperature
self.prompt_model = prompt_model if prompt_model is not None else dspy.settings.lm
self.task_model = task_model if task_model is not None else dspy.settings.lm
self.verbose = verbose
self.track_stats = track_stats
self.teacher_settings = teacher_settings
self.view_data_batch_size = view_data_batch_size
def _print_full_program(self, program):
for i,predictor in enumerate(program.predictors()):
if self.verbose: print(f"Predictor {i}")
if (hasattr(predictor, 'extended_signature')):
if self.verbose: print(f"i: {predictor.extended_signature.instructions}")
if self.verbose: print(f"p: {predictor.extended_signature.fields[-1].name}")
else:
if self.verbose: print(f"i: {predictor.extended_signature1.instructions}")
if self.verbose: print(f"p: {predictor.extended_signature1.fields[-1].name}")
if self.verbose: print("\n")
def _print_model_history(self, model, n=1):
if self.verbose: print(f"Model ({model}) History:")
model.inspect_history(n=n)
def _observe_data(self, trainset):
upper_lim = min(len(trainset), self.view_data_batch_size)
observation = dspy.Predict(DatasetDescriptor, n=1, temperature=1.0)(examples=(trainset[0:upper_lim].__repr__()))
observations = observation["observations"]
skips = 0
for b in range(self.view_data_batch_size, len(trainset), self.view_data_batch_size):
upper_lim = min(len(trainset), b+self.view_data_batch_size)
output = dspy.Predict(DatasetDescriptorWithPriorObservations, n=1, temperature=1.0)(prior_observations=observations, examples=(trainset[b:upper_lim].__repr__()))
if len(output["observations"]) >= 8 and output["observations"][:8].upper() == "COMPLETE":
skips += 1
if skips >= 5:
break
continue
observations += output["observations"]
summary = dspy.Predict(ObservationSummarizer, n=1, temperature=1.0)(observations=observations)
return summary.summary
def _create_example_string(self, fields, example):
# Building the output string
output = []
for field in fields:
name = field.name
separator = field.separator
input_variable = field.input_variable
# Determine the value from input_data or prediction_data
value = example.get(input_variable)
# Construct the string for the current field
field_str = f"{name}{separator}{value}"
output.append(field_str)
# Joining all the field strings
return '\n'.join(output)
def _generate_first_N_candidates(self, module, N, view_data, view_examples, demo_candidates, devset):
candidates = {}
evaluated_candidates = defaultdict(dict)
if view_data:
# Create data observations
self.observations = None
with dspy.settings.context(lm=self.prompt_model):
self.observations = self._observe_data(devset).replace("Observations:","").replace("Summary:","")
if view_examples:
example_sets = {}
for predictor in module.predictors():
# Get all augmented examples
example_set = {}
all_sets_of_examples = demo_candidates[id(predictor)] # Get all generated sets of examples
for example_set_i, set_of_examples in enumerate(all_sets_of_examples):
if example_set_i != 0: # Skip the no examples case
for example in set_of_examples: # Get each individual example in the set
if "augmented" in example.keys():
if example["augmented"]:
if example_set_i not in example_set:
example_set[example_set_i] = []
fields_to_use = predictor.signature.fields
input_variable_names = [field.input_variable for field in fields_to_use]
example_with_only_signature_fields = {key: value for key, value in example.items() if key in input_variable_names}
example_string = self._create_example_string(fields_to_use, example_with_only_signature_fields)
example_set[example_set_i].append(example_string)
example_sets[id(predictor)] = example_set
else:
example_set[example_set_i] = []
example_sets[id(predictor)] = example_set
# Seed the prompt optimizer zero shot with just the instruction, generate BREADTH new prompts
for predictor in module.predictors():
basic_instruction = None
basic_prefix = None
if (hasattr(predictor, 'extended_signature')):
basic_instruction = predictor.extended_signature.instructions
basic_prefix = predictor.extended_signature.fields[-1].name
else:
basic_instruction = predictor.extended_signature1.instructions
basic_prefix = predictor.extended_signature1.fields[-1].name
with dspy.settings.context(lm=self.prompt_model):
# Data & Examples
if view_data and view_examples:
instruct = None
for i in range(1,self.n):
new_instruct = dspy.Predict(BasicGenerateInstructionWithExamplesAndDataObservations, n=1, temperature=self.init_temperature)(basic_instruction=basic_instruction, observations=self.observations, examples=example_sets[id(predictor)][i])
if not instruct:
instruct = new_instruct
else:
instruct.completions.proposed_instruction.extend(new_instruct.completions.proposed_instruction)
instruct.completions.proposed_prefix_for_output_field.extend(new_instruct.completions.proposed_prefix_for_output_field)
# Just data
elif view_data:
instruct = dspy.Predict(BasicGenerateInstructionWithDataObservations, n=N-1, temperature=self.init_temperature)(basic_instruction=basic_instruction, observations=self.observations)
# Just examples
elif view_examples:
instruct = None
for i in range(1,self.n): # Note: skip over the first example set which is empty
new_instruct = dspy.Predict(BasicGenerateInstructionWithExamples, n=1, temperature=self.init_temperature)(basic_instruction=basic_instruction, examples=example_sets[id(predictor)][i])
if not instruct:
instruct = new_instruct
else:
instruct.completions.proposed_instruction.extend(new_instruct.completions.proposed_instruction)
instruct.completions.proposed_prefix_for_output_field.extend(new_instruct.completions.proposed_prefix_for_output_field)
# Neither
else:
instruct = dspy.Predict(BasicGenerateInstruction, n=N-1, temperature=self.init_temperature)(basic_instruction=basic_instruction)
# Add in our initial prompt as a candidate as well
instruct.completions.proposed_instruction.insert(0, basic_instruction)
instruct.completions.proposed_prefix_for_output_field.insert(0, basic_prefix)
candidates[id(predictor)] = instruct.completions
evaluated_candidates[id(predictor)] = {}
if self.verbose: self._print_model_history(self.prompt_model)
return candidates, evaluated_candidates
def compile(self, student, *, devset, optuna_trials_num, max_bootstrapped_demos, max_labeled_demos, eval_kwargs, seed=42, view_data=True, view_examples=True):
random.seed(seed)
# Set up program and evaluation function
module = student.deepcopy()
evaluate = Evaluate(devset=devset, metric=self.metric, **eval_kwargs)
# Generate N few shot example sets
demo_candidates = {}
for i in range(self.n):
if i == 0: # Story empty set of demos as default for index 0
for module_p in module.predictors():
if id(module_p) not in demo_candidates.keys():
demo_candidates[id(module_p)] = []
demo_candidates[id(module_p)].append([])
else:
if self.verbose: print(f"Creating basic bootstrap: {i}/{self.n-1}")
# Create a new basic bootstrap few - shot program .
rng = random.Random(i)
shuffled_devset = devset[:] # Create a copy of devset
rng.shuffle(shuffled_devset) # Shuffle the copy
tp = BootstrapFewShot(metric = self.metric, max_bootstrapped_demos=max_bootstrapped_demos, max_labeled_demos=max_labeled_demos, teacher_settings=self.teacher_settings)
candidate_program = tp.compile(student=module.deepcopy(), trainset=shuffled_devset)
# Store the candidate demos
for module_p, candidate_p in zip(module.predictors(), candidate_program.predictors()):
if id(module_p) not in demo_candidates.keys():
demo_candidates[id(module_p)] = []
demo_candidates[id(module_p)].append(candidate_p.demos)
# Generate N candidate prompts
instruction_candidates, _ = self._generate_first_N_candidates(module, self.n, view_data, view_examples, demo_candidates, devset)
# Initialize variables to store the best program and its score
best_score = float('-inf')
best_program = None
trial_num = 0
trial_logs = {}
# Define our trial objective
def create_objective(baseline_program, instruction_candidates, demo_candidates, evaluate, devset):
def objective(trial):
nonlocal best_program, best_score, trial_num, trial_logs # Allow access to the outer variables
candidate_program = baseline_program.deepcopy()
# Suggest the instruction to use for our predictor
if self.verbose: print(f"Starting trial num: {trial_num}")
trial_logs[trial_num] = {}
for p_old, p_new in zip(baseline_program.predictors(), candidate_program.predictors()):
# Get instruction candidates for our given predictor
p_instruction_candidates = instruction_candidates[id(p_old)]
p_demo_candidates = demo_candidates[id(p_old)]
# Suggest the index of the instruction candidate to use in our trial
instruction_idx = trial.suggest_categorical(f"{id(p_old)}_predictor_instruction",range(len(p_instruction_candidates)))
demos_idx = trial.suggest_categorical(f"{id(p_old)}_predictor_demos",range(len(p_demo_candidates)))
trial_logs[trial_num][f"{id(p_old)}_predictor_instruction"] = instruction_idx
trial_logs[trial_num][f"{id(p_old)}_predictor_demos"] = demos_idx
# Get the selected instruction candidate
selected_candidate = p_instruction_candidates[instruction_idx]
selected_instruction = selected_candidate.proposed_instruction.strip('"').strip()
selected_prefix = selected_candidate.proposed_prefix_for_output_field.strip('"').strip()
# Use this candidates in our program
p_new.extended_signature.instructions = selected_instruction
p_new.extended_signature.fields[-1] = p_new.extended_signature.fields[-1]._replace(name=selected_prefix)
# Get the selected demos
selected_demos = p_demo_candidates[demos_idx]
# Use these demos in our program
p_new.demos = selected_demos
if self.verbose: print("Evaling the following program:")
self._print_full_program(candidate_program)
trial_logs[trial_num]["program"] = candidate_program
# Evaluate with the new prompts
total_score = 0
batch_size = 100
num_batches = math.ceil(len(devset) / batch_size)
for i in range(num_batches):
start_index = i * batch_size
end_index = min((i + 1) * batch_size, len(devset))
split_dev = devset[start_index:end_index]
split_score = evaluate(candidate_program, devset=split_dev, display_table=0)
if self.verbose: print(f"{i}st split score: {split_score}")
total_score += split_score * len(split_dev)
curr_weighted_avg_score = total_score / min((i+1)*100,len(devset))
if self.verbose: print(f"curr average score: {curr_weighted_avg_score}")
trial.report(curr_weighted_avg_score, i)
# Handle pruning based on the intermediate value.
if trial.should_prune():
if self.verbose: print(f"Optuna decided to prune!")
trial_logs[trial_num]["score"] = curr_weighted_avg_score
trial_logs[trial_num]["pruned"] = True
trial_num += 1
raise optuna.TrialPruned()
if self.verbose: print(f"Fully evaled score: {curr_weighted_avg_score}")
self._print_model_history(self.task_model, n=1)
score = curr_weighted_avg_score
trial_logs[trial_num]["score"] = curr_weighted_avg_score
trial_logs[trial_num]["pruned"] = False
# Update the best program if the current score is better
if score > best_score:
best_score = score
best_program = candidate_program.deepcopy()
trial_num += 1
return score
return objective
# Run the trial
objective_function = create_objective(module, instruction_candidates, demo_candidates, evaluate, devset)
sampler = optuna.samplers.TPESampler(seed=seed)
study = optuna.create_study(direction="maximize", sampler=sampler)
score = study.optimize(objective_function, n_trials=optuna_trials_num)
if best_program is not None and self.track_stats:
best_program.trial_logs = trial_logs
return best_program
|