model card + prompt template

README.md

@@ -1,3 +1,143 @@
 ---
-license: cc-by-nc-sa-4.0
+license: apache-2.0
+base_model: Qwen/Qwen2.5-7B-Instruct
+language:
+- ar
+- en
+pipeline_tag: text-generation
+library_name: transformers
+datasets:
+- QCRI/ProBel
+tags:
+- propaganda-detection
+- persuasion-techniques
+- span-identification
+- explainability
+- lora
 ---
+
+# ProBel-MTL
+
+The bilingual multi-task model from the ProBel paper (Mt-SFT): a single
+Qwen2.5-7B-Instruct fine-tune that handles all five ProBel tasks in both Arabic
+and English — binary propaganda detection, coarse-category and fine-grained
+technique classification (with explanations), and technique-labeled span
+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/`.
+
+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).
+
+## Test scores
+
+| | Binary | Coarse | Technique | Span-tag | Span-occ |
+|---|---|---|---|---|---|
+| Arabic  | 0.763 | 0.682 | 0.575 | 0.411 | 0.362 |
+| English | 0.735 | 0.410 | 0.272 | 0.189 | 0.241 |
+
+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
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+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`).
+
+## Instruction format
+
+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):
+
+```
+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...]
+
+Respond EXACTLY in this format (in English):
+Label: true
+Explanation: <your explanation of why this text is or is not propagandistic>
+
+OR
+
+Label: false
+Explanation: <your explanation>
+
+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.
+
+Text: "{TEXT}"
+```
+
+Output formats per task:
+
+| 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 |
+
+Arabic inputs use Arabic prompt variants of the same structure and answer in
+Arabic.
+
+## Intended use and limitations
+
+Built for research on propaganda and persuasion-technique analysis in news and
+social-media text. Predictions are imperfect, technique performance follows the
+long-tailed label distribution (rare techniques are often missed), and outputs
+should support trained human reviewers rather than replace them, particularly
+in moderation or policy settings.
+
+## Citation
+
+```bibtex
+@article{hasanain2026probel,
+  title  = {ProBel: Propaganda Detection with Techniques, Spans, and Explanations},
+  author = {Hasanain, Maram and Hasan, Md Arid and Kmainasi, Mohamed Bayan and
+            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.}
+}
+```

binary_prompt_en.txt

@@ -0,0 +1,31 @@
+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
+- Exaggeration-Minimisation: inflating or downplaying the significance of facts
+- Appeal_to_Fear-Prejudice: exploiting fears or prejudices
+- Causal_Oversimplification: reducing complex issues to simple cause-and-effect
+- Flag_Waving: exploiting national or patriotic sentiments
+- Questioning_the_Reputation: attacking credibility without evidence
+- Doubt: raising questions without providing solid evidence
+- Appeal_to_Authority: citing authority figures to bolster claims
+- Slogans: using catchy, brief phrases to simplify complex issues
+- And other techniques that aim to manipulate the audience's opinion
+
+Respond EXACTLY in this format (in English):
+Label: true
+Explanation: <your explanation of why this text is or is not propagandistic>
+
+OR
+
+Label: false
+Explanation: <your explanation>
+
+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.
+
+Text: "{TEXT}"