src/deepeval/instruction_following_task.py

from datetime import datetime
from src.deepeval.base_task import BaseTask
from deepeval.metrics import PromptAlignmentMetric
from deepeval.test_case import LLMTestCase
from typing import Any

class InstructionFollowingTask(BaseTask):

    def __init__(self, model_name: str):
        super().__init__("metunlp/instruction_following_tr", model_name=model_name)

    def load_dataset_from_hf(self):
        dataset = super().load_dataset_lmjudge_from_hf()
        return dataset

    def evaluate(self) -> dict[str, Any]:
        results = []
        total_model_time = 0
        total_judge_time = 0
        
        for i, row in enumerate(self.dataset):
            start_model = datetime.now()
            input_text = row.get("input", "")
            instruction_text = row.get("instruction", "")

            prompt = (
                f"Girdi: {input_text}\n"
                f"Talimat: {instruction_text}\n"
                f"Çıkıt:"
            )

            output = self.generate_response(prompt, max_new_tokens=200)
            end_model = datetime.now()
            total_model_time += (end_model - start_model).total_seconds()

            start_judge = datetime.now()
            test_case = LLMTestCase(
                input=input_text,
                actual_output=output
            )

            metric = PromptAlignmentMetric(
                prompt_instructions=[instruction_text],
                model="gpt-4o-mini",
                include_reason=True
            )
            metric.measure(test_case)
            end_judge = datetime.now()
            total_judge_time += (end_judge - start_judge).total_seconds()

            results.append({
                "index": i,
                "score": metric.score,
                "reason": metric.reason,
                "score_breakdown": metric.score_breakdown,
                "input": input_text,
                "instruction": instruction_text,
                "output": output
            })
            #Sum all scores in results and divide to nubmer of results
            overallScore = (sum([result["score"] for result in results]) / len(results)) * 100 
        
        print(f"Total model time: {total_model_time} seconds")
        print(f"Total judge time: {total_judge_time} seconds")
        return {"results": overallScore}