Upload 9 files

src/deepeval/bias.py

@@ -0,0 +1,98 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import ast
+
+
+class BiasTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/sosyoloji_bias", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(1, len(dataset))))
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 2
+
+            # Get values from row
+
+            ## common
+            choices = ast.literal_eval(row["choices"])  # Convert string to list
+            formatted_choices = "\n".join([f"{chr(65 + i)}: {choice}" for i, choice in enumerate(choices)])
+
+            ## ambiguous context
+            ambiguous_context = row["ambiguous_context"]
+            ambiguous_question = row["question_ambiguous"]
+            ambiguous_answer = row["answer_ambiguous"]
+            ambiguous_correct_answer_letter = chr(64 + ambiguous_answer)  # 65 - 1 since we need the indexing to start from 0
+            ambiguous_prompt = f"Bağlam: {ambiguous_context}\nSoru: {ambiguous_question}\nSeçenekler:{formatted_choices}"
+
+            ## disambiguated context
+            disambiguated_context = row["disambiguated_context"]
+            disambiguated_question = row["question_disambiguated"]
+            disambiguated_answer = row["answer_disambiguated"]
+            disambiguated_correct_answer_letter = chr(64 + disambiguated_answer)  # 65 - 1 since we need the indexing to start from 0
+            disambiguated_prompt = f"Bağlam: {disambiguated_context}\nSoru: {disambiguated_question}\nSeçenekler:{formatted_choices}"
+
+            # Prints for debugging
+            print(f"Choices: {choices}")
+            print("Type of choices:", type(choices))
+
+            # STAGE 1
+            instruction = ""
+            prompt = f"Stage1:\n{ambiguous_prompt}\n{instruction}\n"
+            message = prompt
+            ambiguous_model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+            ambiguous_model_answer_cleaned = ambiguous_model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+            ## Print answers
+            print(f"Correct Answer: {ambiguous_correct_answer_letter}")
+            print(f"Model Answer: {ambiguous_model_answer}")
+            print(f"Model Answer Cleaned: {ambiguous_model_answer_cleaned}")
+            print(f"Result: {ambiguous_correct_answer_letter == ambiguous_model_answer_cleaned}")
+            ## Check if correct based on metric
+            if ambiguous_correct_answer_letter == ambiguous_model_answer_cleaned:
+                true += 1
+                difficulty_results["ambiguous"]['correct'] += 1
+
+            difficulty_results["ambiguous"]['total'] += 1
+
+            # STAGE 2
+            instruction = ""
+            prompt = f"Stage2:\n{disambiguated_prompt}\n{instruction}\n"
+            message = prompt
+            disambiguated_model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+            disambiguated_model_answer_cleaned = disambiguated_model_answer.strip().replace('\n', '').replace(' ','').upper().replace(':', '')
+            ## Print answers
+            print(f"Correct Answer: {disambiguated_correct_answer_letter}")
+            print(f"Model Answer: {disambiguated_model_answer}")
+            print(f"Model Answer Cleaned: {disambiguated_model_answer_cleaned}")
+            print(f"Result: {disambiguated_correct_answer_letter == disambiguated_model_answer_cleaned}")
+            responses.append((ambiguous_model_answer_cleaned,disambiguated_model_answer_cleaned))
+
+            ## Check if correct based on metric
+            if disambiguated_correct_answer_letter == disambiguated_model_answer_cleaned:
+                true += 1
+                difficulty_results["disambiguated"]['correct'] += 1
+
+            difficulty_results["disambiguated"]['total'] += 1
+
+        # Print results categorized by difficulty
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/math.py

@@ -0,0 +1,128 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import re
+
+class MathTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/math_tr", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(1, len(dataset))))
+
+    def generate_response_oeqa_multi_token(self, msg,max_new_tokens: int = 128):
+        """
+        Handles multiple-choice questions where answers might have multiple tokens.
+        """
+        # Ensure tokenizer has proper special tokens set
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        if self.model.config.pad_token_id is None:
+            self.model.config.pad_token_id = self.tokenizer.pad_token_id
+
+        chat = [
+            {"role": "user", "content": "You are a question-answering chatbot."},
+            {"role": "assistant", "content": "I am ready to answer your questions. Feel free to ask anything.\n"},
+            {"role": "user", "content": f"{msg}"},
+        ]
+        formatted_chat = self.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+        print(formatted_chat)
+
+        inputs = self.tokenizer(formatted_chat, return_tensors="pt", padding=True, truncation=True)
+        input_ids = inputs.input_ids.to(self.model.device)
+        attention_mask = inputs.attention_mask.to(self.model.device)
+
+        # Generate response with proper token limits
+        output = self.model.generate(
+            input_ids,
+            do_sample=True,
+            attention_mask=attention_mask,
+            eos_token_id=self.tokenizer.eos_token_id,
+            pad_token_id=self.tokenizer.pad_token_id,
+            temperature=0.4,
+            max_new_tokens=max_new_tokens,
+        )
+
+        generated_ids = output[0]  # The generated sequence including the prompt
+        generated_tokens = generated_ids[len(input_ids[0]):]  # Exclude the input_ids part
+        generated_text = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+
+        return generated_text
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            category = str(row["difficulty"])
+            answer = row["final_answer"]
+
+            # Prints for debugging
+            print(f"Answer: {answer}")
+            print("Type of answer:", type(answer))
+
+            # Construct the prompt/message
+            instruction = f"""Aşağıdaki matematik problemini verilen nihai cevap formatına uygun olacak şekilde çözün. Tüm adımları gösterdikten sonra, nihai cevabınızı sadece bir kez ve aşağıdaki kurallara uygun şekilde kutu (\\boxed{{}}) içinde verin.
+
+Nihai Cevap için Uyulması Gereken Format Kuralları:
+
+1. Kesirler her zaman en sade hallerinde verilmeli.
+    - Matris içi kesirler: x/y biçiminde.
+    - Diğer tüm kesirler: \\frac{{x}}{{y}} biçiminde.
+2. Çarpma işareti (*) kullanılmamalı. Örnek: 2x yazın, 2**x* değil.
+3. Birden çok değişken varsa alfabetik sıraya uyulmalı ve (x, y, z...), polinomları azalan derece sırasına göre yazılmalı.
+4. Her zaman aynı gösterim biçimi kullanılmalı. Ondalık yerine kesir kullanılmalı (ör. 0.5 yerine \\frac{{1}}{{2}} ).
+5. Faktörize polinomlar daima aynı faktör sırası ile verilsin; her sorguda aynı cevabı verecek şekilde tutarlılığı koruyun.
+6. Nihai cevabı kutu dışında tekrar etmeyin, biçimi değiştirmeyin. Aynı soru tekrarlandığında aynı formatı ve cevabı verin.
+7. Nihai cevap, tek seferde \\boxed{{...}} içinde verilmeli. Örnek: Cevap x ise, "\\boxed{{x}}".
+
+
+Görev: Problemi çözün, son adımda yukarıdaki kurallara tam uyan tek bir kutu içinde nihai cevabı verin.
+
+
+Çözüm:
+
+
+Nihai cevap:
+                        """
+            prompt = f"{instruction}\n\nSoru:\n{row["question"]}\n"
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_oeqa_multi_token(message)
+            responses.append(model_answer)
+            model_answer_cleaned = re.search(r"\\boxed{([^}]*)}", model_answer)
+
+            # Print answers
+            print(f"Correct Answer: {answer}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {answer == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if answer == model_answer_cleaned:
+                true += 1
+                difficulty_results[category]['correct'] += 1
+
+            difficulty_results[category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/metaphors_and_idioms.py

@@ -0,0 +1,87 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import os
+import ast
+import re
+from datasets import load_dataset,get_dataset_split_names
+HF_TOKEN=os.getenv("HF_TOKEN")
+
+class MetaphorsAndIdiomsTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/metaphors_and_idioms", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset # dataset.select(range(min(10, len(dataset))))
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results =  defaultdict(lambda:  defaultdict(lambda: {'correct': 0, 'total': 0}))
+
+        total_count = 0
+        true = 0
+
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            category = "hard" if row["level"]== 1 else "easy" if row["level"] == 0 else None
+            answer_index = row["answer"]
+            correct_answer_letter = chr(65 + answer_index)
+            context = row["context"]
+            choices = ast.literal_eval(row["choices"])  # Convert string to list
+            formatted_choices = "\n".join([f"{chr(65 + i)}: {choice}" for i, choice in enumerate(choices)])
+            subset = row["idiom_type"]
+
+            if subset == "atasözü":
+                question = "Aşağıda verilen durum hangi atasözü ile en iyi ifade edilebilir?"
+            elif subset == "deyim":
+                question = """Verilen bağlamda "[MASKED]" ile boş bırakılan yere hangi deyim getirilirse cümlenin akışı anlamlı olur?"""
+            else:
+                question = "Aşağıda verilen durum hangi atasözü ile en iyi ifade edilebilir?"
+
+            # Prints for debugging
+            print(f"Difficulty: {category}")
+            print("Type of difficulty:", type(category))
+            print(f"Answer: {correct_answer_letter}")
+            print("Type of answer:", type(answer_index))
+
+            # Construct the prompt/message
+            instruction = ""
+            prompt = f"Soru: {question}\nBağlam: {context}\nSeçenekler:\n{formatted_choices}\n{instruction}\n"
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+            responses.append(model_answer)
+            model_answer_cleaned = model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+
+            # Print answers
+            print(f"Correct Answer: {correct_answer_letter}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {correct_answer_letter == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if correct_answer_letter == model_answer_cleaned:
+                true += 1
+                difficulty_results[subset][category]['correct'] += 1
+
+            difficulty_results[subset][category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for subset in difficulty_results.keys():
+            subset_results = difficulty_results[subset]
+            for category, stats in subset_results.items():
+                calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+                print(f"{subset.capitalize()} {category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/mmlu.py

@@ -0,0 +1,87 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import os
+import ast
+import re
+from datasets import load_dataset,get_dataset_config_names
+HF_TOKEN=os.getenv("HF_TOKEN")
+
+class MMLUTask(BaseTask):
+    def __init__(self, model_name):
+        self.subsets = get_dataset_config_names("metunlp/mmlu_tr")
+        print(self.subsets)
+        super().__init__("metunlp/mmlu_tr", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        evaluate_count = 1
+        print("Loading dataset from Hugging Face.")
+        dataset_dict = {}
+        for subset in self.subsets:
+            subset_data = load_dataset(self.dataset_repo, subset, token=HF_TOKEN, split="train")
+            dataset_dict[subset] = subset_data.select(range(min(evaluate_count, len(subset_data))))
+        print("Dataset loaded.")
+        return dataset_dict
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+
+        total_count = 0
+        true = 0
+
+        for subset in self.subsets:
+            curr_dataset = self.dataset[subset]
+            print(curr_dataset[0])
+
+            for row in curr_dataset:
+                total_count += 1
+
+                # Get values from row
+                question = row["question"]
+                answer_index = row["answer"]
+                correct_answer_letter = chr(65 + answer_index)
+                choices = ast.literal_eval(row["choices"])  # Convert string to list
+                formatted_choices = "\n".join([f"{chr(65 + i)}: {choice}" for i, choice in enumerate(choices)])
+
+
+                # Prints for debugging
+                print(f"Answer: {correct_answer_letter}")
+                print("Type of answer:", type(answer_index))
+
+                # Construct the prompt/message
+                instruction = f"Aşağıda {row["subject"]} konusunda çoktan seçmeli bir soru verilmiştir."
+                prompt = f"{instruction}\n\nSoru: {question}\nSeçenekler:\n{formatted_choices}\n\n"
+                message = prompt
+
+                # Get/format answer of the model
+                model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+                responses.append(model_answer)
+                model_answer_cleaned = model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+
+                # Print answers
+                print(f"Correct Answer: {correct_answer_letter}")
+                print(f"Model Answer: {model_answer}")
+                print(f"Model Answer Cleaned: {model_answer_cleaned}")
+                print(f"Result: {correct_answer_letter == model_answer_cleaned}")
+
+                # Check if correct based on metric
+                if correct_answer_letter == model_answer_cleaned:
+                    true += 1
+                    difficulty_results[subset]['correct'] += 1
+
+                difficulty_results[subset]['total'] += 1
+
+        # Print results categorized by subset
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{subset.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/ner.py

@@ -0,0 +1,166 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import re
+
+class NERTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/tr_ner", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(1, len(dataset))))
+
+    def generate_response_oeqa_multi_token(self, msg,max_new_tokens: int = 128):
+        """
+        Handles multiple-choice questions where answers might have multiple tokens.
+        """
+        # Ensure tokenizer has proper special tokens set
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        if self.model.config.pad_token_id is None:
+            self.model.config.pad_token_id = self.tokenizer.pad_token_id
+
+        chat = [
+            {"role": "user", "content": "You are a question-answering chatbot."},
+            {"role": "assistant", "content": "I am ready to answer your questions. Feel free to ask anything.\n"},
+            {"role": "user", "content": f"{msg}"},
+        ]
+        formatted_chat = self.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+        print(formatted_chat)
+
+        inputs = self.tokenizer(formatted_chat, return_tensors="pt", padding=True, truncation=True)
+        input_ids = inputs.input_ids.to(self.model.device)
+        attention_mask = inputs.attention_mask.to(self.model.device)
+
+
+        # Generate response with proper token limits
+        output = self.model.generate(
+            input_ids,
+            do_sample=True,
+            attention_mask=attention_mask,
+            eos_token_id=self.tokenizer.eos_token_id,
+            pad_token_id=self.tokenizer.pad_token_id,
+            temperature=0.4,
+            max_new_tokens=max_new_tokens,
+        )
+
+        generated_ids = output[0]  # The generated sequence including the prompt
+        generated_tokens = generated_ids[len(input_ids[0]):]  # Exclude the input_ids part
+        generated_text = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+
+        return generated_text
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            category = str(row["difficulty"])
+            answer = row["final_answer"]
+
+            # Prints for debugging
+            print(f"Answer: {answer}")
+            print("Type of answer:", type(answer))
+
+            # Construct the prompt/message
+            instruction = ("Aşağıdaki Named Entity Recognition (NER) için etiketlenmesi gereken cümleler vardır. "
+                  "Cümlelerdeki varlıkları belirleyin ve şu kategorilere ayırın: CARDINAL, DATE, EVENT, FAC, GPE, LANGUAGE, LAW, LOC, MONEY, NORP, ORDINAL, ORG, PER, PERCENT, PERSON, PRODUCT, QUANTITY, TIME, TITLE, WORK_OF_ART. "
+                  ""
+                  "Varlıklar, anlamlı bilgiler içeren terimlerdir ve aşağıdaki şekilde tanımlanır: "
+                  "CARDINAL: Nicelik veya sıralama belirtmeyen sayısal ifadeler."
+                  "DATE: Belirli bir tarih veya zaman ifadeleri."
+                  "EVENT: Adlandırılmış olaylar veya durumlar."
+                  "FAC: Binalar veya önemli yerler gibi tesisler."
+                  "GPE: Ülke, şehir veya eyalet gibi coğrafi-politik varlıklar."
+                  "LANGUAGE: Adlandırılmış diller."
+                  "LAW: Yasal belgeler, düzenlemeler veya kanunlar."
+                  "LOC: Coğrafi veya fiziksel konumlar (GPE dışındaki)."
+                  "MONEY: Parasal değerler."
+                  "NORP: Milletler, dini veya siyasi gruplar."
+                  "ORDINAL: Sıralama veya dereceler."
+                  "ORG: Organizasyonlar veya kurumlar."
+                  "PER: Kişisel unvanlar veya sıfatlar."
+                  "PERSON: Bireylerin isimleri."
+                  "PRODUCT: Üretilen nesneler veya araçlar."
+                  "QUANTITY: Ölçülebilir miktarlar ve birimler."
+                  "TIME: Günün belirli saatleri."
+                  "TITLE: Kişi unvanları."
+                  "WORK_OF_ART: Sanat eserleri, kitaplar, müzik vb. Adlar, tarih ifadeleri, konumlar gibi belirgin bilgiler varlıktır."
+                  ""
+                  "Fiiller, sıfatlar, zarflar, soyut kavramlar gibi ifadeler varlık değildir. Çıktıyı aşağıdaki JSON formatında döndürün. "
+                  ""
+                  "Örnekler: "
+                  "Girdi: "
+                  "sentence: \"Üç yıl aradan sonra gerçekleştirilen ve Karadeniz, Ege ve Akdeniz’de düzenlenecek olan tatbikata ilişkin Yunanistan'ın Kathimerini gazetesi 'Türk-Yunan: Çetin donanma dengesinin gücü' başlığını kullandı.\""
+                  "Çıktı: "
+                  "Üç yıl,DATE"
+                  "Karadeniz,LOC"
+                  "Ege,LOC"
+                  "Akdeniz,LOC"
+                  "Yunanistan,GPE"
+                  "Kathimerini,ORG"
+                  "Türk,NORP"
+                  ""
+                  "Girdi:"
+                  "sentence: \"Evlendikten sonra oyunculuğu bırakan Makal, geçen yıl eşi ve oğluyla beraber İstanbul’dan Göcek’e taşınmıştı."
+                  "Çıktı: "
+                  "Makal,PERSON"
+                  "İstanbul,GPE"
+                  "Göcek,GPE"
+                  ""
+                  "Girdi:"
+                  "sentence: \"Yeşil-kırmızılılardan 2016’da ayrılıp 3 sezonluk aradan sonra 2019’da geri dönen Sarıca, takımına 2021 yılında Şampiyonlar Ligi’nde, 2023’te de Süper Lig’de iki final oynattı."
+                  "Çıktı:"
+                  "2016’da,DATE"
+                  "3,CARDINAL"
+                  "2019’da,DATE"
+                  "Sarıca,PERSON"
+                  "2021,DATE"
+                  "Şampiyonlar Ligi’nde,EVENT"
+                  "2023’te,DATE"
+                  "Süper Lig’de,EVENT"
+                  "iki,CARDINAL"
+                  ""
+                  "Verilen cümlelerdeki her varlığı csv formatında yukarıdaki örneklere benzer şekilde belirleyin. Çıktıdaki her satırı aşağıdaki gibi oluşturun: "
+                  "<Varlık metni>,<Varlık etiketi>"),
+            prompt = f"{instruction}\n\nSoru:\n{row["question"]}\n"
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_oeqa_multi_token(message)
+            responses.append(model_answer)
+            model_answer_cleaned = model_answer
+
+            # Print answers
+            print(f"Correct Answer: {answer}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {answer == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if answer == model_answer_cleaned:
+                true += 1
+                difficulty_results[category]['correct'] += 1
+
+            difficulty_results[category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/pos.py

@@ -0,0 +1,159 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import re
+
+class POSTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/tr_pos", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(1, len(dataset))))
+
+    def generate_response_oeqa_multi_token(self, msg,max_new_tokens: int = 128):
+        """
+        Handles multiple-choice questions where answers might have multiple tokens.
+        """
+        # Ensure tokenizer has proper special tokens set
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        if self.model.config.pad_token_id is None:
+            self.model.config.pad_token_id = self.tokenizer.pad_token_id
+
+        chat = [
+            {"role": "user", "content": "You are a question-answering chatbot."},
+            {"role": "assistant", "content": "I am ready to answer your questions. Feel free to ask anything.\n"},
+            {"role": "user", "content": f"{msg}"},
+        ]
+        formatted_chat = self.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+        print(formatted_chat)
+
+        inputs = self.tokenizer(formatted_chat, return_tensors="pt", padding=True, truncation=True)
+        input_ids = inputs.input_ids.to(self.model.device)
+        attention_mask = inputs.attention_mask.to(self.model.device)
+        prompt = ("Aşağıdaki Named Entity Recognition (NER) için etiketlenmesi gereken cümleler vardır. "
+                  "Cümlelerdeki varlıkları belirleyin ve şu kategorilere ayırın: CARDINAL, DATE, EVENT, FAC, GPE, LANGUAGE, LAW, LOC, MONEY, NORP, ORDINAL, ORG, PER, PERCENT, PERSON, PRODUCT, QUANTITY, TIME, TITLE, WORK_OF_ART. "
+                  ""
+                  "Varlıklar, anlamlı bilgiler içeren terimlerdir ve aşağıdaki şekilde tanımlanır: "
+                  "CARDINAL: Nicelik veya sıralama belirtmeyen sayısal ifadeler."
+                  "DATE: Belirli bir tarih veya zaman ifadeleri."
+                  "EVENT: Adlandırılmış olaylar veya durumlar."
+                  "FAC: Binalar veya önemli yerler gibi tesisler."
+                  "GPE: Ülke, şehir veya eyalet gibi coğrafi-politik varlıklar."
+                  "LANGUAGE: Adlandırılmış diller."
+                  "LAW: Yasal belgeler, düzenlemeler veya kanunlar."
+                  "LOC: Coğrafi veya fiziksel konumlar (GPE dışındaki)."
+                  "MONEY: Parasal değerler."
+                  "NORP: Milletler, dini veya siyasi gruplar."
+                  "ORDINAL: Sıralama veya dereceler."
+                  "ORG: Organizasyonlar veya kurumlar."
+                  "PER: Kişisel unvanlar veya sıfatlar."
+                  "PERSON: Bireylerin isimleri."
+                  "PRODUCT: Üretilen nesneler veya araçlar."
+                  "QUANTITY: Ölçülebilir miktarlar ve birimler."
+                  "TIME: Günün belirli saatleri."
+                  "TITLE: Kişi unvanları."
+                  "WORK_OF_ART: Sanat eserleri, kitaplar, müzik vb. Adlar, tarih ifadeleri, konumlar gibi belirgin bilgiler varlıktır."
+                  ""
+                  "Fiiller, sıfatlar, zarflar, soyut kavramlar gibi ifadeler varlık değildir. Çıktıyı aşağıdaki JSON formatında döndürün. "
+                  ""
+                  "Örnekler: "
+                  "Girdi: "
+                  "\"sentence\": \"Üç yıl aradan sonra gerçekleştirilen ve Karadeniz, Ege ve Akdeniz’de düzenlenecek olan tatbikata ilişkin Yunanistan'ın Kathimerini gazetesi 'Türk-Yunan: Çetin donanma dengesinin gücü' başlığını kullandı.\""
+                  "Çıktı: "
+                  "Üç yıl: DATE\" }, { \"text\": \"Karadeniz\", \"label\": \"LOC\" }, { \"text\": \"Ege\", \"label\": \"LOC\" }, { \"text\": \"Akdeniz\", \"label\": \"LOC\" }, { \"text\": \"Yunanistan\", \"label\": \"GPE\" }, { \"text\": \"Kathimerini\", \"label\": \"ORG\" }, { \"text\": \"Türk\", \"label\": \"NORP\" }]} Girdi: {\"sentence\": \"Evlendikten sonra oyunculuğu bırakan Makal, geçen yıl eşi ve oğluyla beraber İstanbul’dan Göcek’e taşınmıştı.\"} Çıktı: {\"entities\": [{ \"text\": \"Makal\", \"label\": \"PERSON\" }, { \"text\": \"İstanbul\", \"label\": \"GPE\" }, { \"text\": \"Göcek\", \"label\": \"GPE\" }]} Girdi: {\"sentence\": \"Yeşil-kırmızılılardan 2016’da ayrılıp 3 sezonluk aradan sonra 2019’da geri dönen Sarıca, takımına 2021 yılında Şampiyonlar Ligi’nde, 2023’te de Süper Lig’de iki final oynattı.\"} Çıktı: {\"entities\": [{ \"text\": \"2016’da\", \"label\": \"DATE\" }, { \"text\": \"3\", \"label\": \"CARDINAL\" }, { \"text\": \"2019’da\", \"label\": \"DATE\" }, { \"text\": \"Sarıca\", \"label\": \"PERSON\" }, { \"text\": \"2021\", \"label\": \"DATE\" }, { \"text\": \"Şampiyonlar Ligi’nde\", \"label\": \"EVENT\" }, { \"text\": \"2023’te\", \"label\": \"DATE\" }, { \"text\": \"Süper Lig’de\", \"label\": \"EVENT\" }, { \"text\": \"iki\", \"label\": \"CARDINAL\" }]}. Verilen cümlelerdeki varlıkları JSON formatında yukarıdaki örneklere benzer şekilde belirleyin. Çıktıyı aşağıdaki gibi oluşturun: Girdi Formatı: {\"sentence\": \"<CÜMLE>\"} Çıktı Formatı: {\"entities\": [{ \"text\": \"<Varlık metni>\", \"label\": \"<Varlık etiketi>\" }]}"),
+
+        # Generate response with proper token limits
+        output = self.model.generate(
+            input_ids,
+            do_sample=True,
+            attention_mask=attention_mask,
+            eos_token_id=self.tokenizer.eos_token_id,
+            pad_token_id=self.tokenizer.pad_token_id,
+            temperature=0.4,
+            max_new_tokens=max_new_tokens,
+        )
+
+        generated_ids = output[0]  # The generated sequence including the prompt
+        generated_tokens = generated_ids[len(input_ids[0]):]  # Exclude the input_ids part
+        generated_text = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+
+        return generated_text
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            category = str(row["difficulty"])
+            answer = row["final_answer"]
+
+            # Prints for debugging
+            print(f"Answer: {answer}")
+            print("Type of answer:", type(answer))
+
+            # Construct the prompt/message
+            instruction = f"""Aşağıdaki matematik problemini verilen nihai cevap formatına uygun olacak şekilde çözün. Tüm adımları gösterdikten sonra, nihai cevabınızı sadece bir kez ve aşağıdaki kurallara uygun şekilde kutu (\\boxed{{}}) içinde verin.
+
+Nihai Cevap için Uyulması Gereken Format Kuralları:
+
+1. Kesirler her zaman en sade hallerinde verilmeli.
+    - Matris içi kesirler: x/y biçiminde.
+    - Diğer tüm kesirler: \\frac{{x}}{{y}} biçiminde.
+2. Çarpma işareti (*) kullanılmamalı. Örnek: 2x yazın, 2**x* değil.
+3. Birden çok değişken varsa alfabetik sıraya uyulmalı ve (x, y, z...), polinomları azalan derece sırasına göre yazılmalı.
+4. Her zaman aynı gösterim biçimi kullanılmalı. Ondalık yerine kesir kullanılmalı (ör. 0.5 yerine \\frac{{1}}{{2}} ).
+5. Faktörize polinomlar daima aynı faktör sırası ile verilsin; her sorguda aynı cevabı verecek şekilde tutarlılığı koruyun.
+6. Nihai cevabı kutu dışında tekrar etmeyin, biçimi değiştirmeyin. Aynı soru tekrarlandığında aynı formatı ve cevabı verin.
+7. Nihai cevap, tek seferde \\boxed{{...}} içinde verilmeli. Örnek: Cevap x ise, "\\boxed{{x}}".
+
+
+Görev: Problemi çözün, son adımda yukarıdaki kurallara tam uyan tek bir kutu içinde nihai cevabı verin.
+
+
+Çözüm:
+
+
+Nihai cevap:
+                        """
+            prompt = f"{instruction}\n\nSoru:\n{row["question"]}\n"
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_oeqa_multi_token(message)
+            responses.append(model_answer)
+            model_answer_cleaned = re.search(r"\\boxed{([^}]*)}", model_answer)
+
+            # Print answers
+            print(f"Correct Answer: {answer}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {answer == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if answer == model_answer_cleaned:
+                true += 1
+                difficulty_results[category]['correct'] += 1
+
+            difficulty_results[category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/sts.py

@@ -0,0 +1,131 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import re
+from datasets import load_dataset
+import os
+from dotenv import load_dotenv
+import openai
+from transformers import AutoModelForCausalLM, AutoTokenizer, LogitsProcessorList, LogitsProcessor
+import torch
+from typing import List
+
+class STSTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/sts_tr", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(1, len(dataset))))
+
+    def generate_response_sts_multi_token(self, msg, max_new_tokens=5, choices: list = []):
+        """
+        Handles multiple-choice questions where answers might have multiple tokens.
+        """
+        # Ensure tokenizer has proper special tokens set
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        if self.model.config.pad_token_id is None:
+            self.model.config.pad_token_id = self.tokenizer.pad_token_id
+
+        chat = [
+            {"role": "user",
+             "content": "You are a sentence similarity scoring chatbot. Only respond with one of the given scores: 0, 1, 2, 3, 4, or 5."},
+            {"role": "assistant", "content": "I am ready to answer your questions. Feel free to ask anything.\n"},
+            {"role": "user", "content": f"{msg}"},
+        ]
+        formatted_chat = self.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+        print(formatted_chat)
+        inputs = self.tokenizer(formatted_chat, return_tensors="pt", padding=True, truncation=True)
+        input_ids = inputs.input_ids.to(self.model.device)
+        attention_mask = inputs.attention_mask.to(self.model.device)
+
+        # Generate the sequence of letters starting from 'A'
+        letters = ["0","1","2","3","4","5"]
+        encoded_choices = [self.tokenizer.encode(letter, add_special_tokens=False) for letter in letters]
+        flattened_encoded_choices = [item for sublist in encoded_choices for item in sublist]  # Flatten the list
+        print(flattened_encoded_choices)
+
+        allowed_tokens = flattened_encoded_choices
+        allowed_tokens += self.get_chat_template_tokens()  # Get the special chat tokens
+        allowed_token_ids = set(allowed_tokens)  # Ensure uniqueness
+
+        # Custom LogitsProcessor to restrict generation
+        class RestrictToABCDLogitsProcessor(LogitsProcessor):
+            def __call__(self, input_ids, scores):
+                mask = torch.full_like(scores, float("-inf"))  # Block all tokens
+                mask[:, list(allowed_token_ids)] = scores[:, list(allowed_token_ids)]  # Allow only A, B, C, D tokens
+                return mask
+
+        logits_processor = LogitsProcessorList([RestrictToABCDLogitsProcessor()])
+
+        # Generate response
+        output = self.model.generate(
+            input_ids,
+            do_sample=True,
+            attention_mask=attention_mask,
+            max_new_tokens=max_new_tokens,
+            eos_token_id=self.tokenizer.eos_token_id,
+            pad_token_id=self.tokenizer.pad_token_id,
+            temperature=0.4,
+            logits_processor=logits_processor,
+        )
+        generated_ids = output[0]  # The generated sequence including the prompt
+        generated_tokens = generated_ids[len(input_ids[0]):]  # Exclude the input_ids part
+        generated_text = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+        return generated_text
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = {'correct': 0, 'total': 0}
+
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            answer = row["score"]
+            choices = ["0","1","2","3","4","5"]
+
+            # Prints for debugging
+            print(f"Answer: {answer}")
+            print("Type of answer:", type(answer))
+
+            # Construct the prompt/message
+            instruction = f"Aşağıda verilen iki cümlenin birbirlerine olan anlamsal benzerliğini 0'dan 5'e kadar olan bir tam sayıyla söyleyin."
+            prompt = f"""{instruction}\nCümle 1: {row["sentence_1"]}\nCümle 2: {row["sentence_2"]}\nSadece tek bir tam sayı söyleyin, ek bir kelime ya da sembol kullanmayın."""
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_sts_multi_token(message, max_new_tokens=2)
+            responses.append(model_answer)
+            model_answer_cleaned = model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+
+            # Print answers
+            print(f"Correct Answer: {answer}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {answer == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if answer == model_answer_cleaned:
+                true += 1
+                difficulty_results['correct'] += 1
+
+            difficulty_results['total'] += 1
+
+        # Print results
+            stats = difficulty_results
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/topic_detection.py

@@ -0,0 +1,79 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import ast
+
+
+class TopicDetectionTask(BaseTask):
+    def __init__(self, model_name):
+        super().__init__("metunlp/topic_detection_tr", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        dataset = super().load_dataset_from_hf()
+        return dataset.select(range(min(10, len(dataset))))
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: {'correct': 0, 'total': 0})
+        total_count = 0
+        true = 0
+
+        for row in self.dataset:
+            total_count += 1
+
+            # Get values from row
+            choices = ast.literal_eval(row["choices"]) # Convert string to list
+            formatted_choices = "\n".join([f"{chr(65+i)}: {choice}" for i, choice in enumerate(choices)])
+            category = row["level"].lower().replace(' ','')
+            answer = row["answer"]
+            text = row["text"]
+
+            # Prints for debugging
+            print(f"Choices: {choices}")
+            print("Type of choices:", type(choices))
+            print("Type of answer:", type(answer))
+
+            # Get answer index (starting from 0)
+            if type(answer) == int:
+                answer_index = answer
+            else:
+                answer_index = int(answer)
+            correct_answer_letter = chr(65 + answer_index)
+
+
+            # Construct the prompt/message
+            instruction = "Aşağıdaki metni analiz et ve seçeneklerden bu metnin en olası kategorisini belirle. Temaya ve detaylara dikkat ederek metnin ana fikrini göz önünde bulundurarak soruyu cevapla."
+            prompt = f"{instruction}\n\nMetin:\n{text}\nSeçenekler:\n{formatted_choices}\n\n"
+            message = prompt
+
+            # Get/format answer of the model
+            model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+            responses.append(model_answer)
+            model_answer_cleaned = model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+
+            # Print answers
+            print(f"Correct Answer: {correct_answer_letter}")
+            print(f"Model Answer: {model_answer}")
+            print(f"Model Answer Cleaned: {model_answer_cleaned}")
+            print(f"Result: {correct_answer_letter == model_answer_cleaned}")
+
+            # Check if correct based on metric
+            if correct_answer_letter == model_answer_cleaned:
+                true += 1
+                difficulty_results[category]['correct'] += 1
+
+            difficulty_results[category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for category, stats in difficulty_results.items():
+            calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+            print(f"{category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+

src/deepeval/turkish_vocabulary.py

@@ -0,0 +1,100 @@
+from src.deepeval.base_task import BaseTask
+from collections import defaultdict
+from src.deepeval.utils import accuracy, accuracy_standard_error
+from typing import Any
+import os
+import ast
+import re
+from datasets import load_dataset,get_dataset_split_names
+HF_TOKEN=os.getenv("HF_TOKEN")
+
+class TurkishVocabularyTask(BaseTask):
+    def __init__(self, model_name):
+        self.subsets = ["rare", "loan"]
+        super().__init__("metunlp/turkish_vocabulary", model_name=model_name)
+
+    def load_dataset_from_hf(self):
+        evaluate_count = 1
+        print("Loading dataset from Hugging Face.")
+        dataset_dict = {}
+        for subset in self.subsets:
+            subset_data = load_dataset(self.dataset_repo, subset, token=HF_TOKEN, split="train")
+            dataset_dict[subset] = subset_data.select(range(min(evaluate_count, len(subset_data))))
+        print("Dataset loaded.")
+        return dataset_dict
+
+
+    def evaluate(self) -> dict[str, Any]:
+        responses = []
+        difficulty_results = defaultdict(lambda: defaultdict(lambda: {'correct': 0, 'total': 0}))
+
+        total_count = 0
+        true = 0
+
+        for subset in self.subsets:
+            curr_dataset = self.dataset[subset]
+            print(curr_dataset[0])
+
+            # Determine the question based on the subset
+            if subset == "rare":
+                question = "Verilen kelimenin eş anlamlısı aşağıdakilerden hangisidir?"
+            elif subset == "loan":
+                question = "Verilen kelimenin Türkçe kökenli eş anlamlısı aşağıdakilerden hangisidir?"
+            else:
+                question = "Verilen kelimenin eş anlamlısı aşağıdakilerden hangisidir?"
+
+            for row in curr_dataset:
+                total_count += 1
+
+                # Get values from row
+                category = "hard" if row["level"]== 1 else "easy" if row["level"] == 0 else None
+                answer_index = row["answer"]
+                correct_answer_letter = chr(65 + answer_index)
+                word = row["word"]
+                choices = ast.literal_eval(row["choices"])  # Convert string to list
+                formatted_choices = "\n".join([f"{chr(65 + i)}: {choice}" for i, choice in enumerate(choices)])
+
+
+
+                # Prints for debugging
+                print(f"Difficulty: {category}")
+                print("Type of difficulty:", type(category))
+                print(f"Answer: {correct_answer_letter}")
+                print("Type of answer:", type(answer_index))
+
+                # Construct the prompt/message
+                instruction = ""
+                prompt = f"Soru: {question}\nKelime: {word}\nSeçenekler:\n{formatted_choices}\n{instruction}\n"
+                message = prompt
+
+                # Get/format answer of the model
+                model_answer = self.generate_response_mcqa_multi_token(message, choices=choices, max_new_tokens=2)
+                responses.append(model_answer)
+                model_answer_cleaned = model_answer.strip().replace('\n', '').replace(' ', '').upper().replace(':','')
+
+                # Print answers
+                print(f"Correct Answer: {correct_answer_letter}")
+                print(f"Model Answer: {model_answer}")
+                print(f"Model Answer Cleaned: {model_answer_cleaned}")
+                print(f"Result: {correct_answer_letter == model_answer_cleaned}")
+
+                # Check if correct based on metric
+                if correct_answer_letter == model_answer_cleaned:
+                    true += 1
+                    difficulty_results[subset][category]['correct'] += 1
+
+                difficulty_results[subset][category]['total'] += 1
+
+        # Print results categorized by difficulty
+        for subset in self.subsets:
+            subset_results = difficulty_results[subset]
+            for category, stats in subset_results.items():
+                calculatedAccuracy = stats['correct'] / stats['total'] if stats['total'] > 0 else 0
+                print(f"{subset.capitalize()} {category.capitalize()} Accuracy: {calculatedAccuracy:.2%} ({stats['correct']}/{stats['total']})")
+
+        print("Results:", responses)
+        print("Overall Accuracy:", true / total_count)
+        acc = accuracy(true, total_count)
+        acc_stderr = accuracy_standard_error(acc, total_count)
+        return {"acc": acc, "acc_stderr": acc_stderr}
+