deepakachu5114

def main():
    train_set, test_set = train_test_split(
        decisions_df, test_size=TEST_SET_SIZE, random_state=RANDOM_STATE
    )
    print(f”Total dataset: {len(decisions_df)} samples”)
    print(f”Training on: {len(train_set)} samples”)
    print(f”Validating on: {len(test_set)} samples”)

    result = {”prompt”: [], “accuracy”: []}

deepakachu5114

optimizer.zero_grad()
loss.backward()
optimizer.step()

new_pred = model(question)
print(new_pred.value)
print(system_prompt.get_gradient_text()) # this prints the textual gradients
print(system_prompt) # the system prompt after optimisation

deepakachu5114

# choose a random data point
idx = 4
question_str = decisions_df.iloc[idx].question
answer_str = decisions_df.iloc[idx].answer

question = tg.Variable(question_str, role_description=”football situaion query”)
answer = tg.Variable(answer_str, role_description=”Ground Truth Answer”)

# get the model prediction before 
pred = model(question)

deepakachu5114

system_prompt = tg.Variable(
    value=”“”You are an expert football referee. I will give you a situation in Football, and you must provide the correct official decision or restart.

Your output must exactly match the decision string. Do not provide any extra explanation, just the decision.
“”“,
    requires_grad=True,
    role_description=”System prompt to answer football queries”
)
model = tg.BlackboxLLM(llm_engine, system_prompt)

deepakachu5114

# Install the library
!pip install textgrad -q

# --- Imports ---
import pandas as pd
import textgrad as tg
from textgrad import get_engine, Variable
from textgrad.optimizer import TextualGradientDescent
from sklearn.model_selection import train_test_split