AI? ??? ???? ???? ?? ???? ????? ? ???? ??? ??? ????? ??? ??? ?????. ??? ?? ????? ?? ??? ??? ???? AI? ??? ???? ???? ?? ???? ???? ??? ?? ??? ?????.

? ?????? 139? ?? ??? ??? ??? ???? ?????? ???? ? ??? ?? OpenAI? GPT-4o-mini? ?? ???? ?? DSPy? ??? ??? ???? ??? ???????. ???? ?? ???? ?????? ?? ?? DSPy? ???? ?? AI ??? ???? ??, ?? ? ????? ??? ?? ???.

---

????? ??

??, ??? ?? ? LLM ???? ??? ????? AI? ??? ??? ??? ??? ??? ??? ??? ?????? ??? ?????. ????? DSPy ? ?? ?? ????? ?????? ?? Python 3? ????? ?????.

pip install dspy datasets

????? ??? ? 2GB??? ??? ??? ?? ? ??? ? ? ?? ?? ? ????.

?? ? ???? ?? ??? ??? ???? ???? ???? ??? ??? ????.

from datasets import load_dataset

# Load the dataset
ds = load_dataset("artem9k/ai-text-detection-pile", split="train")

# For simplicity, we’ll do an even split of data for testing and training
NUM_EXAMPLES = 80  # Adjust this to experiment with dataset size
ds = ds.train_test_split(test_size=NUM_EXAMPLES, train_size=NUM_EXAMPLES)

?: NUM_EXAMPLES?? ???? ? ? ?????? ????? ??? ?? ? ??? ??? ? ????.

---

?? ??

???? OpenAI? GPT-4o-mini? ???? ?? DSPy ???? ???????. GPT-4o-mini? OpenAI? GPT-4o ??? ?? ???? ??? ?? ??????. DSPy? ???? ??-?? ??? ???? ??? ???? ? ????? ??????.

??? ???? ?? "YOUR_API_KEY"? OpenAI API ?? ????.

import dspy
from typing import Literal

# Initialize the OpenAI GPT-4o-mini model
lm = dspy.LM('openai/gpt-4o-mini', api_key="YOUR_API_KEY")
dspy.configure(lm=lm, experimental=True)

# Define the AI text detector signature
class DetectAiText(dspy.Signature):
    """Classify text as written by human or by AI."""
    text: str = dspy.InputField()
    source: Literal['ai', 'human'] = dspy.OutputField()

# Create a basic predictor
detector = dspy.Predict(DetectAiText)

????? ???? ?????? ???? ?????. ?? ??? DSPy? ???? ???? ??? ??? ?????.

? ?? ?? ???? "???"? ???? ? ????.

print(detector(text="Hello world (this definitely wasn't written by AI)"))

??? ??? '??' ??? ?????.

---

??? ??

?? ?? ???? ???? DSPy? ?? ??? ???? ??? ??? ?????. ?? ?? ??? ??? ??(?? ?? AI)? ???? ????? ???? ??? ????? ???????.

??? ???? ???? ??? ??? ????.

from dspy.evaluate import Evaluate

# Define a simple evaluation metric
def validate_text_source(example: dspy.Example, pred, trace=None) -> int:
    return 1 if example.source.lower() == pred.source.lower() else 0

# Transform the dataset into DSPy-compatible "Example" objects
dspy_trainset = [
    dspy.Example(source=x['source'], text=x['text']).with_inputs('text') for x in ds['train']
]
dspy_devset = [
    dspy.Example(source=x['source'], text=x['text']).with_inputs('text') for x in ds['test']
]

# Evaluate the detector
evaluator = Evaluate(devset=dspy_devset, num_threads=12)  # Adjust threads based on your system
evaluator(detector, metric=validate_text_source)

?? ????? 76%~81%? ???? ??????. ??? ??? ??? ????? ?? ??? ??? ? ????.

---

DSPy? ???

DSPy? ??? ?? ??? ??? ????. MIPROv2 ??? ????? ???? ????? ???? ???? ??? ???? ??? ???? ? ????. ?????? ? ?? ??, ?? ??? ? ?? ?? ??? ???? ? ????? ??????.

??? ????? ???? ???? ??? ??? ????.

pip install dspy datasets

??: "light" ?? ??? ??? ?? ??? ?? ??? ????? 80?? ??? ??? ?? 0.50?? ?????.

---

?? ? ??

???? ??? ? ??? ?? ???? ?? ??????. ? ?? ????? ??? 76%~81%? ???? 91.25%? ???? ??????. ?? ??? 81.2% ~ 91.25% ????? ???? ???? ??? ??? ???????.

?? ??? ?? ???? ??? ?????:

from datasets import load_dataset

# Load the dataset
ds = load_dataset("artem9k/ai-text-detection-pile", split="train")

# For simplicity, we’ll do an even split of data for testing and training
NUM_EXAMPLES = 80  # Adjust this to experiment with dataset size
ds = ds.train_test_split(test_size=NUM_EXAMPLES, train_size=NUM_EXAMPLES)

??? ?? ??? ??? ? ????.

• ?????? ?? ????(?, ?, ?)? ????? ??? ????? ?? ?????.
• ?? ? ??? ?? ??? ?? ?? ???
• ?? ????? ????? LLM?? ??????.

---

??

? ? ????? DSPy? ?? ?? ??? ?? LLM ???? ????? ??? ??????. ?? ???? ????? ?? AI? ??? ???? ???? ? ?? ?? ??? ??? ??????. ? ??? ????? ??? DSPy? ??? ??? ???? ??? ????? ?? ??? AI ??? ?? ??? ??? ???.

DSPy? ??? ??? ?? ?? ??? ??? LLM ??? ??? ? ?? ?? ?????.

---

GitHub?? ?? ??? ??? ? ????.

??? ?????? ??? ?????. DSPy? ??? ???? ?????!

LinkedIn?? ?? ??? ? ???? | CTO ? ??? @ EES.

? ??? DSPy? ???? AI ?? ???? ????? OpenAI? GPT-mini ???? ?? ?????. ??? ??? PHP ??? ????? ?? ?? ??? ?????!