Replace template README with model-linked card

README.md

@@ -4,14 +4,46 @@ emoji: 💬
 colorFrom: yellow
 colorTo: purple
 sdk: gradio
-sdk_version: 6.5.1
+sdk_version: 6.31.0
 app_file: app.py
 pinned: false
-hf_oauth: true
-hf_oauth_scopes:
-- inference-api
-license: apache-2.0
-short_description: A model to split hadith to sanad and matn
+license: mit
+short_description: Split a raw hadith into sanad and matn
+models:
+  - SHK4K/hadith-segmentation-bert
+language:
+  - ar
+tags:
+  - hadith
+  - sanad
+  - matn
+  - arabic
+  - token-classification
+  - digital-humanities
 ---
 
-An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
+# فاصل الحديث — Hadith Splitter
+
+يقسم نص حديث خام (بدون علامات ترقيم هيكلية) إلى **سند** (سلسلة الرواة) و**متن**
+(قول النبي صلى الله عليه وسلم)، ويعرض الحدود بينهما بالألوان على النص الأصلي.
+
+Splits a raw hadith string into its **sanad** (chain of narrators) and **matn**
+(the saying itself), and highlights the boundary on the original text.
+
+- Model: [SHK4K/hadith-segmentation-bert](https://huggingface.co/SHK4K/hadith-segmentation-bert)
+  (AraBERT v02 fine-tuned for token classification)
+- Labels: `B-SANAD`, `I-SANAD`, `B-MATN`, `I-MATN`, reduced to a single cut
+- Diacritics are stripped for inference (the model was trained undiacritized)
+  and mapped back, so the output keeps your original vocalized text
+- Long hadiths: processed in overlapping 512-token windows (stride 128) with
+  per-word majority voting, so nothing is truncated
+- Output strings are sliced from your original text, so punctuation and spacing
+  are preserved
+
+## Limits
+
+- Accuracy is limited on unusual orthography; the model was trained on
+  normalized Arabic, and the app normalizes before inference.
+- Pastes longer than ~25k subword tokens are processed only up to that point.
+- The model is small (0.1B); treat the boundary as a suggestion on
+  low-resource texts, not ground truth.