Wise harmfulness detection

assets/ar/factuality_disinformation_harmful_content/harmfulness_detection/CT22Harmful_GPT4_FewShot_Arabic.py

@@ -0,0 +1,85 @@
+from llmebench.datasets import CT22HarmfulDataset
+from llmebench.models import OpenAIModel
+from llmebench.tasks import HarmfulDetectionTask
+
+
+def metadata():
+    return {
+        "author": "Mohamed Bayan Kmainasi, Rakif Khan, Ali Ezzat Shahroor, Boushra Bendou, Maram Hasanain, and Firoj Alam",
+        "affiliation": "Arabic Language Technologies, Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University (HBKU)",
+        "model": "GPT-4o-2024-05-22",
+        "description": "For a comprehensive analysis and results, refer to our peer-reviewed publication available at [Springer](https://doi.org/10.1007/978-981-96-0576-7_30) or explore the preprint version on [arXiv](https://arxiv.org/abs/2409.07054).",
+    }
+
+
+def config():
+    return {
+        "dataset": CT22HarmfulDataset,
+        "task": HarmfulDetectionTask,
+        "model": OpenAIModel,
+        "model_args": {
+            "max_tries": 3,
+        },
+        "general_args": {"test_split": "ar", "fewshot": {"train_split": "ar"}},
+    }
+
+
+def prompt(input_sample, examples):
+    base_prompt = "هل قد تتسبب هذه التغريدة بأي ضرر أو أذى لشخص أو مؤسسة؟ إن كانت كذلك، أجب بـ 'ضارة'. أجب بـ 'غير ضارة' إن لم تكن كذلك. قدم فقط التصنيف.\n"
+    prompt = few_shot_prompt(input_sample, base_prompt, examples)
+
+    return [
+        {
+            "role": "system",
+            "content": "أنت خبير في تحليل وتصنيف التغريدات.",
+        },
+        {
+            "role": "user",
+            "content": prompt,
+        },
+    ]
+
+
+def few_shot_prompt(input_sample, base_prompt, examples):
+    out_prompt = base_prompt + "\n"
+    for example in examples:
+        # Found chatgpt confused when using 0 and 1 in the prompt
+        label = "غير ضارة" if example["label"] == "0" else "ضارة"
+        out_prompt = (
+            out_prompt + "التغريدة: " + example["input"] + "التصنيف: " + label + "\n\n"
+        )
+
+    # Append the sentence we want the model to predict for but leave the Label blank
+    out_prompt = out_prompt + "التغريدة: " + input_sample + "التصنيف: \n"
+
+    # print("=========== FS Prompt =============\n")
+    # print(out_prompt)
+
+    return out_prompt
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+    label_fixed = None
+    label = label.replace("label:", "")
+    label = label.replace("التصنيف:", "")
+    label = label.strip().lower()
+
+    if (
+        "غير ضارة" in label
+        or "ليست ضارة" in label
+        or "غير ضاره" in label
+        or "غير" in label
+        or "not" in label
+        or "ليست" in label
+        or "لا" in label
+        or "not harmful" in label
+        or label.startswith("no")
+        or "safe" in label
+        or "not_harmful" in label
+    ):
+        label_fixed = "0"
+    elif "ضارة" in label or "harmful" in label or "نعم" in label or "yes" in label:
+        label_fixed = "1"
+
+    return label_fixed

assets/ar/factuality_disinformation_harmful_content/harmfulness_detection/CT22Harmful_GPT4_FewShot_English.py

@@ -0,0 +1,84 @@
+from llmebench.datasets import CT22HarmfulDataset
+from llmebench.models import OpenAIModel
+from llmebench.tasks import HarmfulDetectionTask
+
+
+def metadata():
+    return {
+        "author": "Mohamed Bayan Kmainasi, Rakif Khan, Ali Ezzat Shahroor, Boushra Bendou, Maram Hasanain, and Firoj Alam",
+        "affiliation": "Arabic Language Technologies, Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University (HBKU)",
+        "model": "GPT-4o-2024-05-22",
+        "description": "For a comprehensive analysis and results, refer to our peer-reviewed publication available at [Springer](https://doi.org/10.1007/978-981-96-0576-7_30) or explore the preprint version on [arXiv](https://arxiv.org/abs/2409.07054).",
+    }
+
+
+def config():
+    return {
+        "dataset": CT22HarmfulDataset,
+        "task": HarmfulDetectionTask,
+        "model": OpenAIModel,
+        "model_args": {
+            "max_tries": 3,
+        },
+        "general_args": {"test_split": "ar", "fewshot": {"train_split": "ar"}},
+    }
+
+
+def prompt(input_sample, examples):
+    base_prompt = "Can this tweet cause harm to any person, organization or entity? Answer only by 'harmful' or 'not harmful'. Provide only label.\n"
+    prompt = few_shot_prompt(input_sample, base_prompt, examples)
+
+    return [
+        {
+            "role": "system",
+            "content": "You are an expert in tweet analysis and classification.",
+        },
+        {
+            "role": "user",
+            "content": prompt,
+        },
+    ]
+
+
+def few_shot_prompt(input_sample, base_prompt, examples):
+    out_prompt = base_prompt + "\n"
+    for example in examples:
+        # Found chatgpt confused when using 0 and 1 in the prompt
+        label = "harmful" if example["label"] == "1" else "not harmful"
+        out_prompt = (
+            out_prompt + "tweet: " + example["input"] + "\nlabel: " + label + "\n\n"
+        )
+
+    # Append the sentence we want the model to predict for but leave the label blank
+    out_prompt = out_prompt + "tweet: " + input_sample + "\nlabel: \n"
+
+    # print("=========== FS Prompt =============\n")
+    # print(out_prompt)
+
+    return out_prompt
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+    label = label.replace("label:", "")
+    label = label.replace("التصنيف:", "")
+    label = label.strip().lower()
+
+    if (
+        "غير ضارة" in label
+        or "ليست ضارة" in label
+        or "غير ضاره" in label
+        or "غير" in label
+        or "not" in label
+        or "ليست" in label
+        or "لا" in label
+        or "not harmful" in label
+        or label.startswith("no")
+        or "safe" in label
+        or "not_harmful" in label
+    ):
+        return "0"
+    elif "ضارة" in label or "harmful" in label or "نعم" in label or "yes" in label:
+        return "1"
+
+    return None

assets/ar/factuality_disinformation_harmful_content/harmfulness_detection/CT22Harmful_GPT4_FewShot_Mixed.py

@@ -0,0 +1,85 @@
+from llmebench.datasets import CT22HarmfulDataset
+from llmebench.models import OpenAIModel
+from llmebench.tasks import HarmfulDetectionTask
+
+
+def metadata():
+    return {
+        "author": "Mohamed Bayan Kmainasi, Rakif Khan, Ali Ezzat Shahroor, Boushra Bendou, Maram Hasanain, and Firoj Alam",
+        "affiliation": "Arabic Language Technologies, Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University (HBKU)",
+        "model": "GPT-4o-2024-05-22",
+        "description": "For a comprehensive analysis and results, refer to our peer-reviewed publication available at [Springer](https://doi.org/10.1007/978-981-96-0576-7_30) or explore the preprint version on [arXiv](https://arxiv.org/abs/2409.07054).",
+    }
+
+
+def config():
+    return {
+        "dataset": CT22HarmfulDataset,
+        "task": HarmfulDetectionTask,
+        "model": OpenAIModel,
+        "model_args": {
+            "max_tries": 3,
+        },
+        "general_args": {"test_split": "ar", "fewshot": {"train_split": "ar"}},
+    }
+
+
+def prompt(input_sample, examples):
+    base_prompt = "هل قد تتسبب هذه التغريدة بأي ضرر أو أذى لشخص أو مؤسسة؟ إن كانت كذلك، أجب بـ 'harmful'. أجب بـ 'not harmful' إن لم تكن كذلك. قدم فقط التصنيف.\n"
+    prompt = few_shot_prompt(input_sample, base_prompt, examples)
+
+    return [
+        {
+            "role": "system",
+            "content": "أنت خبير في تحليل وتصنيف التغريدات.",
+        },
+        {
+            "role": "user",
+            "content": prompt,
+        },
+    ]
+
+
+def few_shot_prompt(input_sample, base_prompt, examples):
+    out_prompt = base_prompt + "\n"
+    for example in examples:
+        # Found chatgpt confused when using 0 and 1 in the prompt
+        label = "not harmful" if example["label"] == "0" else "harmful"
+        out_prompt = (
+            out_prompt + "التغريدة: " + example["input"] + "التصنيف: " + label + "\n\n"
+        )
+
+    # Append the sentence we want the model to predict for but leave the Label blank
+    out_prompt = out_prompt + "التغريدة: " + input_sample + "التصنيف: \n"
+
+    # print("=========== FS Prompt =============\n")
+    # print(out_prompt)
+
+    return out_prompt
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+    label_fixed = None
+    label = label.replace("label:", "")
+    label = label.replace("التصنيف:", "")
+    label = label.strip().lower()
+
+    if (
+        "غير ضارة" in label
+        or "ليست ضارة" in label
+        or "غير ضاره" in label
+        or "غير" in label
+        or "not" in label
+        or "ليست" in label
+        or "لا" in label
+        or "not harmful" in label
+        or label.startswith("no")
+        or "safe" in label
+        or "not_harmful" in label
+    ):
+        return "0"
+    elif "ضارة" in label or "harmful" in label or "نعم" in label or "yes" in label:
+        return "1"
+
+    return label_fixed

assets/ar/factuality_disinformation_harmful_content/harmfulness_detection/CT22Harmful_GPT4_ZeroShot_Arabic.py

@@ -0,0 +1,71 @@
+import re
+
+from llmebench.datasets import CT22HarmfulDataset
+from llmebench.models import OpenAIModel
+from llmebench.tasks import HarmfulDetectionTask
+
+
+def metadata():
+    return {
+        "author": "Mohamed Bayan Kmainasi, Rakif Khan, Ali Ezzat Shahroor, Boushra Bendou, Maram Hasanain, and Firoj Alam",
+        "affiliation": "Arabic Language Technologies, Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University (HBKU)",
+        "model": "GPT-4o-2024-05-22",
+        "description": "For a comprehensive analysis and results, refer to our peer-reviewed publication available at [Springer](https://doi.org/10.1007/978-981-96-0576-7_30) or explore the preprint version on [arXiv](https://arxiv.org/abs/2409.07054).",
+    }
+
+
+def config():
+    return {
+        "dataset": CT22HarmfulDataset,
+        "task": HarmfulDetectionTask,
+        "model": OpenAIModel,
+        "model_args": {
+            "class_labels": ["0", "1"],
+            "max_tries": 30,
+        },
+        "general_args": {"test_split": "ar"},
+    }
+
+
+def prompt(input_sample):
+    prompt_string = (
+        f"هل قد تتسبب هذه التغريدة بأي ضرر أو أذى لشخص أو مؤسسة؟ إن كانت كذلك، أجب بـ 'ضارة'. أجب بـ 'غير ضارة' إن لم تكن كذلك. قدم فقط التصنيف.\n\n"
+        f"التغريدة: {input_sample}\n"
+        f"التصنيف: \n"
+    )
+    return [
+        {
+            "role": "system",
+            "content": "أنت خبير في تحليل وتصنيف التغريدات.",
+        },
+        {
+            "role": "user",
+            "content": prompt_string,
+        },
+    ]
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+    label = label.replace("label:", "")
+    label = label.replace("التصنيف:", "")
+    label = label.strip().lower()
+
+    if (
+        "غير ضارة" in label
+        or "ليست ضارة" in label
+        or "غير ضاره" in label
+        or "غير" in label
+        or "not" in label
+        or "ليست" in label
+        or "لا" in label
+        or "not harmful" in label
+        or label.startswith("no")
+        or "safe" in label
+        or "not_harmful" in label
+    ):
+        return "0"
+    elif "ضارة" in label or "harmful" in label or "نعم" in label or "yes" in label:
+        return "1"
+
+    return None