Add factuality, disinformation, harmful content assets for JAIS

assets/ar/factuality_disinformation_harmful_content/adult_content_detection/Adult_JAIS13b_ZeroShot.py

@@ -0,0 +1,46 @@
+from llmebench.datasets import AdultDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import AdultTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": AdultDataset,
+        "task": AdultTask,
+        "model": FastChatModel,
+        "model_args": {
+            "class_labels": ["ADULT", "NOT_ADULT"],
+            "max_tries": 3,
+        },
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        f'Given the following tweet, label it as "ADULT" or "NOT_ADULT" based on the content of the tweet.\n\n'
+        f"tweet: {input_sample}\n"
+        f"label: \n"
+    )
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    out = response["choices"][0]["message"]["content"].replace("label: ", "")
+    j = out.find(".")
+    if j > 0:
+        out = out[0:j]
+    return out

assets/ar/factuality_disinformation_harmful_content/attentionworthy/CT22Attentionworthy_JAIS13b_ZeroShot.py

@@ -0,0 +1,73 @@
+from llmebench.datasets import CT22AttentionworthyDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import AttentionworthyTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": CT22AttentionworthyDataset,
+        "task": AttentionworthyTask,
+        "model": FastChatModel,
+        "model_args": {
+            "class_labels": [
+                "yes_discusses_action_taken",
+                "harmful",
+                "yes_discusses_cure",
+                "yes_asks_question",
+                "no_not_interesting",
+                "yes_other",
+                "yes_blame_authorities",
+                "yes_contains_advice",
+                "yes_calls_for_action",
+            ],
+            "max_tries": 30,
+        },
+        "general_args": {"test_split": "ar"},
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        f'Annotate "tweet" into one of the following categories: yes_discusses_action_taken, harmful, yes_discusses_cure, yes_asks_question, no_not_interesting, yes_other, yes_blame_authorities, yes_contains_advice, yes_calls_for_action\n\n'
+        f"tweet: {input_sample}\n"
+        f"label: \n"
+    )
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+
+    label = (
+        label.lower()
+        .replace(" - ", ", ")
+        .replace(",", "")
+        .replace(".", "")
+        .replace("label:", "")
+    )
+    label = label.strip()
+    # label = re.sub("\s+", "_", label)
+    if label.startswith("no"):
+        label_fixed = "no_not_interesting"
+    elif label == "yes_discusses_covid-19_vaccine_side_effects":
+        label_fixed = "yes_discusses_cure"
+    elif label == "yes_harmful":
+        label_fixed = "harmful"
+    elif label.startswith("yes"):
+        label_fixed = label
+
+    return label_fixed

assets/ar/factuality_disinformation_harmful_content/checkworthyness/CT22Checkworthiness_JAIS13b_ZeroShot.py

@@ -0,0 +1,64 @@
+import re
+
+from llmebench.datasets import CT22CheckworthinessDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import CheckworthinessTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": CT22CheckworthinessDataset,
+        "task": CheckworthinessTask,
+        "model": FastChatModel,
+        "model_args": {
+            "class_labels": ["0", "1"],
+            "max_tries": 30,
+        },
+        "general_args": {"test_split": "ar"},
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        f'Annotate the "tweet" into "one" of the following categories: checkworthy or not_checkworthy\n\n'
+        f"tweet: {input_sample}\n"
+        f"label: \n"
+    )
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+
+    label = label.replace("label:", "").strip()
+
+    if "label: " in label:
+        arr = label.split("label: ")
+        label = arr[1].strip()
+
+    if label == "checkworthy" or label == "Checkworthy":
+        label_fixed = "1"
+    elif label == "Not_checkworthy." or label == "not_checkworthy":
+        label_fixed = "0"
+    elif "not_checkworthy" in label or "label: not_checkworthy" in label:
+        label_fixed = "0"
+    elif "checkworthy" in label or "label: checkworthy" in label:
+        label_fixed = "1"
+    else:
+        label_fixed = None
+
+    return label_fixed

assets/ar/factuality_disinformation_harmful_content/claim_detection/CT22Claim_JAIS13b_ZeroShot.py

@@ -0,0 +1,56 @@
+from llmebench.datasets import CT22ClaimDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import ClaimDetectionTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": CT22ClaimDataset,
+        "task": ClaimDetectionTask,
+        "model": FastChatModel,
+        "model_args": {
+            "class_labels": ["0", "1"],
+            "max_tries": 30,
+        },
+        "general_args": {"test_split": "ar"},
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        f"Given the following tweet, please identify if it contains a claim. If it does, annotate 'yes', if it does not, annotate 'no'\n\n"
+        f"tweet: {input_sample}\n"
+        f"label: \n"
+    )
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+
+    label = label.replace("label:", "").strip()
+
+    if "label: " in label:
+        arr = label.split("label: ")
+        label = arr[1].strip()
+
+    if label == "yes" or label == "the sentence contains a factual claim":
+        label_fixed = "1"
+    if label == "no":
+        label_fixed = "0"
+
+    return label_fixed

assets/ar/factuality_disinformation_harmful_content/factuality/ANSFactuality_JAIS13b_ZeroShot.py

@@ -0,0 +1,63 @@
+from llmebench.datasets import ANSFactualityDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import FactualityTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": ANSFactualityDataset,
+        "task": FactualityTask,
+        "model": FastChatModel,
+        "model_args": {
+            "max_tries": 3,
+        },
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        "Detect whether the information in the sentence is factually true or false. "
+        "Answer only by true or false.\n\n"
+        + "Sentence: "
+        + input_sample
+        + "\nlabel: \n"
+    )
+
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    input_label = response["choices"][0]["message"]["content"]
+    input_label = input_label.replace(".", "").strip().lower()
+
+    if (
+        "true" in input_label
+        or "label: 1" in input_label
+        or "label: yes" in input_label
+    ):
+        pred_label = "true"
+    elif (
+        "false" in input_label
+        or "label: 0" in input_label
+        or "label: no" in input_label
+    ):
+        pred_label = "false"
+    else:
+        print("label problem!! " + input_label)
+        pred_label = None
+
+    return pred_label

assets/ar/factuality_disinformation_harmful_content/factuality/COVID19Factuality_JAIS13b_ZeroShot.py

@@ -0,0 +1,55 @@
+from llmebench.datasets import COVID19FactualityDataset
+from llmebench.models import FastChatModel
+from llmebench.tasks import FactualityTask
+
+
+def metadata():
+    return {
+        "author": "Arabic Language Technologies, QCRI, HBKU",
+        "model": "JAIS-13b",
+        "description": "Locally hosted JAIS-13b-chat model using FastChat.",
+        "scores": {"Macro-F1": ""},
+    }
+
+
+def config():
+    return {
+        "dataset": COVID19FactualityDataset,
+        "task": FactualityTask,
+        "model": FastChatModel,
+        "model_args": {
+            "class_labels": ["yes", "no"],
+            "max_tries": 30,
+        },
+    }
+
+
+def prompt(input_sample):
+    base_prompt = (
+        f'Annotate the "tweet" into one of the following categories: correct or incorrect\n\n'
+        f"tweet: {input_sample}\n"
+        f"label: \n"
+    )
+    return [
+        {
+            "role": "user",
+            "content": base_prompt,
+        },
+    ]
+
+
+def post_process(response):
+    label = response["choices"][0]["message"]["content"]
+
+    if label.startswith("I am unable to verify".lower()) or label.startswith(
+        "I am unable to categorize".lower()
+    ):
+        label_fixed = None
+    elif "label: incorrect" in label or "incorrect" in label:
+        label_fixed = "no"
+    elif "label: correct" in label or "correct" in label:
+        label_fixed = "yes"
+    else:
+        label_fixed = None
+
+    return label_fixed