prompt templates for all ten language-task pairs; card with verified examples

README.md

@@ -27,13 +27,13 @@ extraction in two output formats.
 Trained with LoRA (r=16, alpha=32) on the Arabic and English training splits of
 [QCRI/ProBel](https://huggingface.co/datasets/QCRI/ProBel) across all five task
 formats jointly; the checkpoint was selected on validation loss and merged into
-the base model, so it loads as a regular causal LM. The LoRA adapter alone is in
-`lora_adapter/`.
+the base model, so it loads as a regular causal LM. The LoRA adapter alone is
+in `lora_adapter/`.
 
 Companion resources: [dataset](https://huggingface.co/datasets/QCRI/ProBel) ·
 [code](https://github.com/MohamedBayan/ProBel) · paper: *ProBel: Propaganda
-Detection with Techniques, Spans, and Explanations* (PDF in the code
-repository; arXiv link will be added once available).
+Detection with Techniques, Spans, and Explanations* (arXiv preprint; the link
+will be added here once the listing is live).
 
 ## Test scores
 
@@ -46,80 +46,81 @@ Binary is macro-F1; coarse/technique are micro-F1; spans use the
 overlap-adjusted micro-F1 of Da San Martino et al. (2020). These match the
 paper's Mt-SFT rows and were produced with greedy decoding.
 
-## Usage
+## Prompt templates
 
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+The model expects the exact task prompts it was trained on.
+[`prompts/templates.json`](prompts/templates.json) ships all ten of them —
+`{arabic, english} x {binary, coarse, multilabel, span_tag, span_match_occ}` —
+each a `{"system": ..., "user": ...}` pair where the user message contains a
+`{TEXT}` placeholder for the input sentence.
 
-model = AutoModelForCausalLM.from_pretrained("QCRI/ProBel-MTL",
-                                             torch_dtype="bfloat16",
-                                             device_map="auto")
-tok = AutoTokenizer.from_pretrained("QCRI/ProBel-MTL")
-
-system = ("You are an expert in media analysis and propaganda detection, "
-          "specialized in English text. You analyze texts to determine whether "
-          "they contain propaganda techniques and provide clear, evidence-based "
-          "explanations for your assessments.")
-user = open("binary_prompt_en.txt").read().replace(
-    "{TEXT}", "Even sadder, however, is the fact that these smear campaigns "
-              "work most of the time.")
-
-ids = tok.apply_chat_template(
-    [{"role": "system", "content": system}, {"role": "user", "content": user}],
-    add_generation_prompt=True, return_tensors="pt").to(model.device)
-out = model.generate(ids, max_new_tokens=512, do_sample=False)
-print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
-# Label: false
-# Explanation: ...
-```
-
-The model also serves directly with vLLM (`vllm serve QCRI/ProBel-MTL`).
+| Task | Model output |
+|---|---|
+| `binary` | `Label: true` or `Label: false`, then `Explanation: ...` |
+| `coarse` / `multilabel` | `Labels: <names or none>`, then `Explanation: ...` |
+| `span_tag` | the input sentence with inline `<span type="Technique">...</span>` tags |
+| `span_match_occ` | a JSON list of `{"text", "label", "occurrence"}` objects |
 
-## Instruction format
+Arabic templates carry the same task instructions with an Arabic-specialized
+system prompt; the model answers Arabic inputs in Arabic.
 
-The model expects the exact task prompts it was trained on. The full templates
-for all five tasks and both languages ship with the
-[dataset conversion scripts](https://github.com/MohamedBayan/ProBel)
-(`src/ms_swift/prepare_data.py`); the binary English prompt looks like this
-(`{TEXT}` is the input sentence):
+## Usage
 
+```bash
+pip install "transformers>=4.51" accelerate
 ```
-Analyze the following text and determine whether it contains propaganda techniques.
 
-Propaganda techniques include manipulative language strategies such as:
-- Loaded_Language: emotionally charged words to influence perception
-- Name_Calling-Labeling: attaching negative labels to dismiss someone
-[...technique glossary...]
+Binary detection with an explanation:
 
-Respond EXACTLY in this format (in English):
-Label: true
-Explanation: <your explanation of why this text is or is not propagandistic>
+```python
+import json
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
-OR
+templates = json.load(open(hf_hub_download("QCRI/ProBel-MTL", "prompts/templates.json")))
+model = AutoModelForCausalLM.from_pretrained("QCRI/ProBel-MTL",
+                                             dtype="bfloat16", device_map="auto")
+tok = AutoTokenizer.from_pretrained("QCRI/ProBel-MTL")
 
-Label: false
-Explanation: <your explanation>
+def run(task, lang, text, max_new_tokens=512):
+    t = templates[lang][task]
+    msgs = [{"role": "system", "content": t["system"]},
+            {"role": "user", "content": t["user"].replace("{TEXT}", text)}]
+    ids = tok.apply_chat_template(msgs, add_generation_prompt=True,
+                                  return_tensors="pt").to(model.device)
+    out = model.generate(ids, max_new_tokens=max_new_tokens, do_sample=False)
+    return tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True)
+
+print(run("binary", "english",
+          "The corrupt elites are destroying everything we hold dear."))
+# Label: true
+# Explanation: The paragraph relies on a sweeping, emotive accusation that
+# unnamed "elites" are ruining "everything we hold dear" ...
+```
 
-Notes:
-- Use "true" if the text contains ANY propaganda technique, "false" otherwise.
-- The Label MUST be exactly "true" or "false" (lowercase).
-- The Explanation should identify which specific techniques are used (if any) and why.
-- Analyze the text objectively, considering the language, framing, and intent.
+Technique-labeled span extraction (same helper):
 
-Text: "{TEXT}"
+```python
+print(run("span_tag", "english",
+          "The corrupt elites are destroying everything we hold dear."))
+# <span type="Appeal_to_Fear-Prejudice">The corrupt elites are destroying
+# everything we hold dear.</span>
+
+print(run("multilabel", "arabic",
+          "الإعلام الكاذب يواصل نشر أكاذيبه المسمومة لتضليل الشعب."))
+# Labels: Loaded_Language, Questioning_the_Reputation
+# Explanation: يستخدم النص لغة محملة بالعواطف مثل "الكاذب" و"أكاذيبه المسمومة" ...
 ```
 
-Output formats per task:
+The model also serves directly with vLLM:
 
-| Task | Output |
-|---|---|
-| binary | `Label: true\|false` + `Explanation: ...` |
-| coarse / technique | `Labels: <comma-separated or none>` + `Explanation: ...` |
-| span-tag | the input sentence with inline `<span type="Technique">...</span>` tags |
-| span-occ | a JSON list of `{"text", "label", "occurrence"}` objects |
+```bash
+vllm serve QCRI/ProBel-MTL
+```
 
-Arabic inputs use Arabic prompt variants of the same structure and answer in
-Arabic.
+The parsers that turn the span outputs back into character offsets, and the
+full evaluation pipeline, are in the
+[code repository](https://github.com/MohamedBayan/ProBel).
 
 ## Intended use and limitations
 
@@ -138,6 +139,6 @@ in moderation or policy settings.
             Sartori, Elisa and Shahroor, Ali Ezzat and Da San Martino, Giovanni
             and Alam, Firoj},
   year   = {2026},
-  note   = {Preprint. This entry will be updated upon publication.}
+  note   = {arXiv preprint}
 }
 ```

prompts/templates.json

@@ -0,0 +1,46 @@
+{
+  "arabic": {
+    "binary": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in Arabic text. You analyze texts to determine whether they contain propaganda techniques and provide clear, evidence-based explanations for your assessments.",
+      "user": "Analyze the following text and determine whether it contains propaganda techniques.\n\nPropaganda techniques include manipulative language strategies such as:\n- Loaded_Language: emotionally charged words to influence perception\n- Name_Calling-Labeling: attaching negative labels to dismiss someone\n- Exaggeration-Minimisation: inflating or downplaying the significance of facts\n- Appeal_to_Fear-Prejudice: exploiting fears or prejudices\n- Causal_Oversimplification: reducing complex issues to simple cause-and-effect\n- Flag_Waving: exploiting national or patriotic sentiments\n- Questioning_the_Reputation: attacking credibility without evidence\n- Doubt: raising questions without providing solid evidence\n- Appeal_to_Authority: citing authority figures to bolster claims\n- Slogans: using catchy, brief phrases to simplify complex issues\n- And other techniques that aim to manipulate the audience's opinion\n\nRespond EXACTLY in this format (in English):\nLabel: true\nExplanation: <your explanation of why this text is or is not propagandistic>\n\nOR\n\nLabel: false\nExplanation: <your explanation>\n\nNotes:\n- Use \"true\" if the text contains ANY propaganda technique, \"false\" otherwise.\n- The Label MUST be exactly \"true\" or \"false\" (lowercase).\n- The Explanation should identify which specific techniques are used (if any) and why.\n- Analyze the text objectively, considering the language, framing, and intent.\n\nText: \"{TEXT}\""
+    },
+    "coarse": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in Arabic text. You categorize propaganda techniques into broad categories and provide clear, evidence-based explanations.",
+      "user": "Analyze the following text and identify which broad propaganda categories are present.\n\nPropaganda Categories (with descriptions and included techniques):\n\n- Manipulative_Wording: The text uses emotionally charged, exaggerated, or loaded language to influence the reader's perception. This includes: Loaded_Language (emotionally charged words), Exaggeration-Minimisation (inflating/downplaying facts), Obfuscation-Vagueness-Confusion (deliberately vague language).\n\n- Reputation: The text attacks, questions, or manipulates the reputation or credibility of a person, group, or institution. This includes: Name_Calling-Labeling (attaching negative labels), Questioning_the_Reputation (undermining credibility), Guilt_by_Association (linking to negative entities), Doubt (raising unfounded questions).\n\n- Justification: The text uses appeals to authority, emotions, values, or popularity to justify a position without proper evidence. This includes: Appeal_to_Authority (citing experts), Appeal_to_Fear-Prejudice (exploiting fears), Appeal_to_Pity (exploiting sympathy), Appeal_to_Popularity (bandwagon), Appeal_to_Values (exploiting shared values), Flag_Waving (patriotic sentiments).\n\n- Simplification: The text oversimplifies complex issues, causes, or consequences to push a particular narrative. This includes: Causal_Oversimplification (simplifying causes), Consequential_Oversimplification (simplifying outcomes), False_Dilemma-No_Choice (only two options presented).\n\n- Call: The text uses slogans, conversation killers, or urgency to push the audience toward action or shut down critical thinking. This includes: Slogans (catchy brief phrases), Conversation_Killer (shutting down discussion), Appeal_to_Time (false urgency).\n\n- Distraction: The text diverts attention from the main argument using irrelevant topics, counter-accusations, or misrepresentations. This includes: Red_Herring (irrelevant diversions), Whataboutism (counter-accusations), Straw_Man (misrepresenting arguments), Appeal_to_Hypocrisy (tu quoque fallacy), Repetition (repeated messaging).\n\nRespond EXACTLY in this format (in English):\nLabels: Category1, Category2\nExplanation: <your explanation of why each category applies>\n\nIf the text contains NO propaganda:\nLabels: none\nExplanation: <your explanation of why this text is not propagandistic>\n\nNotes:\n- List ALL categories that apply, separated by commas.\n- Use the EXACT category names: Manipulative_Wording, Reputation, Justification, Simplification, Call, Distraction.\n- Multiple categories can apply to the same text.\n- The Explanation should identify which specific techniques within each category are present.\n- Analyze the text objectively, considering the language, framing, and rhetorical strategies used.\n\nText: \"{TEXT}\""
+    },
+    "multilabel": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in Arabic text. You identify specific propaganda techniques used in texts and provide detailed, evidence-based explanations.",
+      "user": "Analyze the following text and identify ALL propaganda techniques present.\n\nPropaganda Techniques (with descriptions):\n- Appeal_to_Authority: Citing an authority figure or expert to bolster a claim, even when the cited authority lacks relevant expertise.\n- Appeal_to_Fear-Prejudice: Exploiting existing fears, anxieties, or prejudices to promote a conclusion rather than using rational arguments.\n- Appeal_to_Hypocrisy: Deflecting criticism by accusing the opponent of the same behavior, rather than addressing the argument.\n- Appeal_to_Pity: Exploiting the audience's sympathy or compassion to gain support, bypassing logical reasoning.\n- Appeal_to_Popularity: Arguing something is true because many people believe it, regardless of evidence.\n- Appeal_to_Time: Creating a false sense of urgency to push the audience toward a decision without careful consideration.\n- Appeal_to_Values: Exploiting deeply held values (religious, cultural, moral) to emotionally sway the audience.\n- Causal_Oversimplification: Reducing a complex issue with multiple causes to a single simplistic cause-and-effect.\n- Consequential_Oversimplification: Oversimplifying the consequences of an action, presenting only one possible outcome.\n- Conversation_Killer: Using cliches or authoritative statements that shut down further discussion or critical thinking.\n- Doubt: Raising questions about credibility without providing solid evidence, creating uncertainty.\n- Exaggeration-Minimisation: Inflating or downplaying the importance of something to distort perception.\n- False_Dilemma-No_Choice: Presenting only two options as if they are the only possibilities.\n- Flag_Waving: Exploiting national, patriotic, or group identity feelings to justify an action.\n- Guilt_by_Association: Discrediting someone by associating them with something negative without logical connection.\n- Loaded_Language: Using emotionally charged words that carry strong connotations beyond their literal meaning.\n- Name_Calling-Labeling: Attaching negative labels to dismiss or discredit without engaging with substance.\n- Obfuscation-Vagueness-Confusion: Using vague or complex language to confuse the audience or obscure the message.\n- Questioning_the_Reputation: Attacking credibility or reputation to undermine arguments or authority.\n- Red_Herring: Introducing an irrelevant topic to divert attention from the main issue.\n- Repetition: Deliberately repeating a message to make it more memorable and persuasive.\n- Slogans: Using brief, catchy phrases to simplify complex issues and rally support.\n- Straw_Man: Misrepresenting someone's argument to make it easier to attack.\n- Whataboutism: Responding to criticism with a counter-accusation about a different issue.\n\nRespond EXACTLY in this format (in English):\nLabels: Technique1, Technique2, Technique3\nExplanation: <your explanation of why each technique applies>\n\nIf the text contains NO propaganda:\nLabels: none\nExplanation: <your explanation of why this text is not propagandistic>\n\nNotes:\n- List ALL techniques that apply, separated by commas.\n- Use the EXACT technique names as listed above (e.g., \"Loaded_Language\", not \"loaded language\").\n- Multiple techniques can apply to the same text.\n- The Explanation should reference specific parts of the text that exemplify each technique.\n- Analyze the text objectively, considering the language, framing, and rhetorical strategies used.\n\nText: \"{TEXT}\""
+    },
+    "span_tag": {
+      "system": "You are a precise propaganda detection model specialized in Arabic text analysis. You identify propaganda techniques at the span level, extracting exact text spans and classifying them.",
+      "user": "Identify propaganda techniques in the given Arabic text.\nCopy the ENTIRE input text and surround propaganda spans with XML tags.\n\nTag Format: <span type=\"TECHNIQUE_NAME\">text</span>\n\nTechniques:\n  - Loaded_Language: كلمات أو عبارات مشحونة عاطفياً تهدف للتأثير على القارئ\n  - Name_Calling-Labeling: إلصاق تسميات أو ألقاب سلبية بالأشخاص أو الكيانات\n  - Exaggeration-Minimisation: تضخيم أو تقليل أهمية الحقائق والأحداث\n  - Questioning_the_Reputation: التشكيك في سمعة أو مصداقية شخص أو جهة\n  - Obfuscation-Vagueness-Confusion: استخدام لغة غامضة أو مبهمة لإرباك القارئ\n  - Causal_Oversimplification: تبسيط مفرط للأسباب والعلاقات السببية\n  - Doubt: إثارة الشكوك حول شيء ما دون تقديم أدلة\n  - Appeal_to_Authority: الاستشهاد بسلطة أو خبير لتعزيز الحجة\n  - Flag_Waving: استخدام المشاعر الوطنية أو القومية للتأثير\n  - Repetition: التكرار المتعمد لكلمات أو عبارات لترسيخها في الأذهان\n  - Slogans: استخدام شعارات جذابة ومختصرة\n  - Appeal_to_Fear-Prejudice: استغلال المخاوف أو التحيزات\n  - Consequential_Oversimplification: تبسيط مفرط للنتائج والعواقب\n  - Appeal_to_Hypocrisy: اتهام الطرف الآخر بالنفاق لتحويل الانتباه\n  - False_Dilemma-No_Choice: تقديم خيارين فقط كأنهما الخياران الوحيدان\n  - Appeal_to_Time: استخدام ضغط الوقت أو الإلحاح\n  - Conversation_Killer: عبارات تهدف لإنهاء النقاش\n  - Appeal_to_Values: استغلال القيم المشتركة للتأثير\n  - Red_Herring: تحويل الانتباه عن الموضوع الأصلي\n  - Appeal_to_Popularity: الاحتكام إلى رأي الأغلبية\n  - Appeal_to_Pity: استغلال التعاطف أو الشفقة للتأثير على الرأي\n  - Guilt_by_Association: الربط بين شخص وجهة سلبية\n  - Straw_Man: تشويه حجة الطرف الآخر لتسهيل مهاجمتها\n  - Whataboutism: الرد على انتقاد بانتقاد مضاد بدلاً من الإجابة\n\nNotes:\n- Your output MUST include a copy of the entire input text, including non-tagged parts.\n- If no propaganda is found, copy the input text exactly without any tags.\n- Spans may overlap. For overlapping spans, use separate tags (nesting is allowed).\n- Do not alter the text in any way. Do not output explanations.\n- Start generating output straight away.\n\nText: {TEXT}"
+    },
+    "span_match_occ": {
+      "system": "You are a precise propaganda detection model specialized in Arabic text analysis. You identify propaganda techniques at the span level, extracting exact text spans and classifying them.",
+      "user": "Identify propaganda techniques in the given Arabic text.\nFor each technique found, extract the exact text span and classify it.\n\nOutput Format:\n[{\"text\": \"exact span from input\", \"label\": \"TECHNIQUE_NAME\", \"occurrence\": N}]\n\nTechniques:\n  - Loaded_Language: كلمات أو عبارات مشحونة عاطفياً تهدف للتأثير على القارئ\n  - Name_Calling-Labeling: إلصاق تسميات أو ألقاب سلبية بالأشخاص أو الكيانات\n  - Exaggeration-Minimisation: تضخيم أو تقليل أهمية الحقائق والأحداث\n  - Questioning_the_Reputation: التشكيك في سمعة أو مصداقية شخص أو جهة\n  - Obfuscation-Vagueness-Confusion: استخدام لغة غامضة أو مبهمة لإرباك القارئ\n  - Causal_Oversimplification: تبسيط مفرط للأسباب والعلاقات السببية\n  - Doubt: إثارة الشكوك حول شيء ما دون تقديم أدلة\n  - Appeal_to_Authority: الاستشهاد بسلطة أو خبير لتعزيز الحجة\n  - Flag_Waving: استخدام المشاعر الوطنية أو القومية للتأثير\n  - Repetition: التكرار المتعمد لكلمات أو عبارات لترسيخها في الأذهان\n  - Slogans: استخدام شعارات جذابة ومختصرة\n  - Appeal_to_Fear-Prejudice: استغلال المخاوف أو التحيزات\n  - Consequential_Oversimplification: تبسيط مفرط للنتائج والعواقب\n  - Appeal_to_Hypocrisy: اتهام الطرف الآخر بالنفاق لتحويل الانتباه\n  - False_Dilemma-No_Choice: تقديم خيارين فقط كأنهما الخياران الوحيدان\n  - Appeal_to_Time: استخدام ضغط الوقت أو الإلحاح\n  - Conversation_Killer: عبارات تهدف لإنهاء النقاش\n  - Appeal_to_Values: استغلال القيم المشتركة للتأثير\n  - Red_Herring: تحويل الانتباه عن الموضوع الأصلي\n  - Appeal_to_Popularity: الاحتكام إلى رأي الأغلبية\n  - Appeal_to_Pity: استغلال التعاطف أو الشفقة للتأثير على الرأي\n  - Guilt_by_Association: الربط بين شخص وجهة سلبية\n  - Straw_Man: تشويه حجة الطرف الآخر لتسهيل مهاجمتها\n  - Whataboutism: الرد على انتقاد بانتقاد مضاد بدلاً من الإجابة\n\nNotes:\n- Copy spans EXACTLY from the input text, character-for-character.\n- Spans may overlap: the same text region can have multiple techniques.\n- Use \"occurrence\" to distinguish repeated text (1=first, 2=second, etc.).\n- If no propaganda is found, return an empty array: []\n- Return a valid JSON array only. Do not output explanations.\n\nText: {TEXT}\n\nOutput:"
+    }
+  },
+  "english": {
+    "binary": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in English text. You analyze texts to determine whether they contain propaganda techniques and provide clear, evidence-based explanations for your assessments.",
+      "user": "Analyze the following text and determine whether it contains propaganda techniques.\n\nPropaganda techniques include manipulative language strategies such as:\n- Loaded_Language: emotionally charged words to influence perception\n- Name_Calling-Labeling: attaching negative labels to dismiss someone\n- Exaggeration-Minimisation: inflating or downplaying the significance of facts\n- Appeal_to_Fear-Prejudice: exploiting fears or prejudices\n- Causal_Oversimplification: reducing complex issues to simple cause-and-effect\n- Flag_Waving: exploiting national or patriotic sentiments\n- Questioning_the_Reputation: attacking credibility without evidence\n- Doubt: raising questions without providing solid evidence\n- Appeal_to_Authority: citing authority figures to bolster claims\n- Slogans: using catchy, brief phrases to simplify complex issues\n- And other techniques that aim to manipulate the audience's opinion\n\nRespond EXACTLY in this format (in English):\nLabel: true\nExplanation: <your explanation of why this text is or is not propagandistic>\n\nOR\n\nLabel: false\nExplanation: <your explanation>\n\nNotes:\n- Use \"true\" if the text contains ANY propaganda technique, \"false\" otherwise.\n- The Label MUST be exactly \"true\" or \"false\" (lowercase).\n- The Explanation should identify which specific techniques are used (if any) and why.\n- Analyze the text objectively, considering the language, framing, and intent.\n\nText: \"{TEXT}\""
+    },
+    "coarse": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in English text. You categorize propaganda techniques into broad categories and provide clear, evidence-based explanations.",
+      "user": "Analyze the following text and identify which broad propaganda categories are present.\n\nPropaganda Categories (with descriptions and included techniques):\n\n- Manipulative_Wording: The text uses emotionally charged, exaggerated, or loaded language to influence the reader's perception. This includes: Loaded_Language (emotionally charged words), Exaggeration-Minimisation (inflating/downplaying facts), Obfuscation-Vagueness-Confusion (deliberately vague language).\n\n- Reputation: The text attacks, questions, or manipulates the reputation or credibility of a person, group, or institution. This includes: Name_Calling-Labeling (attaching negative labels), Questioning_the_Reputation (undermining credibility), Guilt_by_Association (linking to negative entities), Doubt (raising unfounded questions).\n\n- Justification: The text uses appeals to authority, emotions, values, or popularity to justify a position without proper evidence. This includes: Appeal_to_Authority (citing experts), Appeal_to_Fear-Prejudice (exploiting fears), Appeal_to_Popularity (bandwagon), Appeal_to_Values (exploiting shared values), Flag_Waving (patriotic sentiments).\n\n- Simplification: The text oversimplifies complex issues, causes, or consequences to push a particular narrative. This includes: Causal_Oversimplification (simplifying causes), Consequential_Oversimplification (simplifying outcomes), False_Dilemma-No_Choice (only two options presented).\n\n- Call: The text uses slogans, conversation killers, or urgency to push the audience toward action or shut down critical thinking. This includes: Slogans (catchy brief phrases), Conversation_Killer (shutting down discussion), Appeal_to_Time (false urgency).\n\n- Distraction: The text diverts attention from the main argument using irrelevant topics, counter-accusations, or misrepresentations. This includes: Red_Herring (irrelevant diversions), Whataboutism (counter-accusations), Straw_Man (misrepresenting arguments), Appeal_to_Hypocrisy (tu quoque fallacy), Repetition (repeated messaging).\n\nRespond EXACTLY in this format (in English):\nLabels: Category1, Category2\nExplanation: <your explanation of why each category applies>\n\nIf the text contains NO propaganda:\nLabels: none\nExplanation: <your explanation of why this text is not propagandistic>\n\nNotes:\n- List ALL categories that apply, separated by commas.\n- Use the EXACT category names: Manipulative_Wording, Reputation, Justification, Simplification, Call, Distraction.\n- Multiple categories can apply to the same text.\n- The Explanation should identify which specific techniques within each category are present.\n- Analyze the text objectively, considering the language, framing, and rhetorical strategies used.\n\nText: \"{TEXT}\""
+    },
+    "multilabel": {
+      "system": "You are an expert in media analysis and propaganda detection, specialized in English text. You identify specific propaganda techniques used in texts and provide detailed, evidence-based explanations.",
+      "user": "Analyze the following text and identify ALL propaganda techniques present.\n\nPropaganda Techniques (with descriptions):\n- Appeal_to_Authority: Citing an authority figure or expert to bolster a claim, even when the cited authority lacks relevant expertise.\n- Appeal_to_Fear-Prejudice: Exploiting existing fears, anxieties, or prejudices to promote a conclusion rather than using rational arguments.\n- Appeal_to_Hypocrisy: Deflecting criticism by accusing the opponent of the same behavior, rather than addressing the argument.\n- Appeal_to_Popularity: Arguing something is true because many people believe it, regardless of evidence.\n- Appeal_to_Time: Creating a false sense of urgency to push the audience toward a decision without careful consideration.\n- Appeal_to_Values: Exploiting deeply held values (religious, cultural, moral) to emotionally sway the audience.\n- Causal_Oversimplification: Reducing a complex issue with multiple causes to a single simplistic cause-and-effect.\n- Consequential_Oversimplification: Oversimplifying the consequences of an action, presenting only one possible outcome.\n- Conversation_Killer: Using cliches or authoritative statements that shut down further discussion or critical thinking.\n- Doubt: Raising questions about credibility without providing solid evidence, creating uncertainty.\n- Exaggeration-Minimisation: Inflating or downplaying the importance of something to distort perception.\n- False_Dilemma-No_Choice: Presenting only two options as if they are the only possibilities.\n- Flag_Waving: Exploiting national, patriotic, or group identity feelings to justify an action.\n- Guilt_by_Association: Discrediting someone by associating them with something negative without logical connection.\n- Loaded_Language: Using emotionally charged words that carry strong connotations beyond their literal meaning.\n- Name_Calling-Labeling: Attaching negative labels to dismiss or discredit without engaging with substance.\n- Obfuscation-Vagueness-Confusion: Using vague or complex language to confuse the audience or obscure the message.\n- Questioning_the_Reputation: Attacking credibility or reputation to undermine arguments or authority.\n- Red_Herring: Introducing an irrelevant topic to divert attention from the main issue.\n- Repetition: Deliberately repeating a message to make it more memorable and persuasive.\n- Slogans: Using brief, catchy phrases to simplify complex issues and rally support.\n- Straw_Man: Misrepresenting someone's argument to make it easier to attack.\n- Whataboutism: Responding to criticism with a counter-accusation about a different issue.\n\nRespond EXACTLY in this format (in English):\nLabels: Technique1, Technique2, Technique3\nExplanation: <your explanation of why each technique applies>\n\nIf the text contains NO propaganda:\nLabels: none\nExplanation: <your explanation of why this text is not propagandistic>\n\nNotes:\n- List ALL techniques that apply, separated by commas.\n- Use the EXACT technique names as listed above (e.g., \"Loaded_Language\", not \"loaded language\").\n- Multiple techniques can apply to the same text.\n- The Explanation should reference specific parts of the text that exemplify each technique.\n- Analyze the text objectively, considering the language, framing, and rhetorical strategies used.\n\nText: \"{TEXT}\""
+    },
+    "span_tag": {
+      "system": "You are a precise propaganda detection model specialized in English text analysis. You identify propaganda techniques at the span level, extracting exact text spans and classifying them.",
+      "user": "Identify propaganda techniques in the given English text.\nCopy the ENTIRE input text and surround propaganda spans with XML tags.\n\nTag Format: <span type=\"TECHNIQUE_NAME\">text</span>\n\nTechniques:\n  - Loaded_Language: emotionally charged words or phrases used to influence the reader\n  - Name_Calling-Labeling: attaching negative labels to people or entities\n  - Exaggeration-Minimisation: inflating or downplaying the significance of facts\n  - Questioning_the_Reputation: undermining the credibility of a person or entity\n  - Obfuscation-Vagueness-Confusion: using vague or ambiguous language to confuse\n  - Causal_Oversimplification: oversimplifying causes and causal relationships\n  - Doubt: raising doubts without providing evidence\n  - Appeal_to_Authority: citing authority figures or experts to strengthen argument\n  - Flag_Waving: exploiting national or patriotic sentiments\n  - Repetition: deliberate repetition of words or phrases for emphasis\n  - Slogans: use of catchy, brief phrases\n  - Appeal_to_Fear-Prejudice: exploiting fears or prejudices\n  - Consequential_Oversimplification: oversimplifying consequences\n  - Appeal_to_Hypocrisy: accusing the opponent of hypocrisy to deflect\n  - False_Dilemma-No_Choice: presenting only two options as if they are the only ones\n  - Appeal_to_Time: using urgency or time pressure\n  - Conversation_Killer: phrases designed to end discussion\n  - Appeal_to_Values: exploiting shared values to influence\n  - Red_Herring: diverting attention from the main topic\n  - Appeal_to_Popularity: appealing to majority opinion\n  - Guilt_by_Association: linking a person to a negative entity\n  - Straw_Man: distorting the opponent's argument to make it easier to attack\n  - Whataboutism: responding to criticism with counter-criticism instead of answering\n\nNotes:\n- Your output MUST include a copy of the entire input text, including non-tagged parts.\n- If no propaganda is found, copy the input text exactly without any tags.\n- Spans may overlap. For overlapping spans, use separate tags (nesting is allowed).\n- Do not alter the text in any way. Do not output explanations.\n- Start generating output straight away.\n\nText: {TEXT}"
+    },
+    "span_match_occ": {
+      "system": "You are a precise propaganda detection model specialized in English text analysis. You identify propaganda techniques at the span level, extracting exact text spans and classifying them.",
+      "user": "Identify propaganda techniques in the given English text.\nFor each technique found, extract the exact text span and classify it.\n\nOutput Format:\n[{\"text\": \"exact span from input\", \"label\": \"TECHNIQUE_NAME\", \"occurrence\": N}]\n\nTechniques:\n  - Loaded_Language: emotionally charged words or phrases used to influence the reader\n  - Name_Calling-Labeling: attaching negative labels to people or entities\n  - Exaggeration-Minimisation: inflating or downplaying the significance of facts\n  - Questioning_the_Reputation: undermining the credibility of a person or entity\n  - Obfuscation-Vagueness-Confusion: using vague or ambiguous language to confuse\n  - Causal_Oversimplification: oversimplifying causes and causal relationships\n  - Doubt: raising doubts without providing evidence\n  - Appeal_to_Authority: citing authority figures or experts to strengthen argument\n  - Flag_Waving: exploiting national or patriotic sentiments\n  - Repetition: deliberate repetition of words or phrases for emphasis\n  - Slogans: use of catchy, brief phrases\n  - Appeal_to_Fear-Prejudice: exploiting fears or prejudices\n  - Consequential_Oversimplification: oversimplifying consequences\n  - Appeal_to_Hypocrisy: accusing the opponent of hypocrisy to deflect\n  - False_Dilemma-No_Choice: presenting only two options as if they are the only ones\n  - Appeal_to_Time: using urgency or time pressure\n  - Conversation_Killer: phrases designed to end discussion\n  - Appeal_to_Values: exploiting shared values to influence\n  - Red_Herring: diverting attention from the main topic\n  - Appeal_to_Popularity: appealing to majority opinion\n  - Guilt_by_Association: linking a person to a negative entity\n  - Straw_Man: distorting the opponent's argument to make it easier to attack\n  - Whataboutism: responding to criticism with counter-criticism instead of answering\n\nNotes:\n- Copy spans EXACTLY from the input text, character-for-character.\n- Spans may overlap: the same text region can have multiple techniques.\n- Use \"occurrence\" to distinguish repeated text (1=first, 2=second, etc.).\n- If no propaganda is found, return an empty array: []\n- Return a valid JSON array only. Do not output explanations.\n\nText: {TEXT}\n\nOutput:"
+    }
+  }
+}