Upload app.py with huggingface_hub

app.py

@@ -14,16 +14,20 @@ from fastapi.staticfiles import StaticFiles
 from pydantic import BaseModel
 from transformers import AutoTokenizer
 
-from routing_model import MAX_ROUTES, RouterModel, build_text, route
+from routing_model import MAX_ROUTES, RouterModel, build_text, detect_lang, route
 
 HF_TOKEN = os.environ.get("MODEL_HF_TOKEN") or os.environ.get("HF_TOKEN")
 
-# Two backbones on the same head and the same task. The 6M is a bidirectional
-# BERT encoder; the 52M is a Llama decoder. Measured on the same eval sets in one
-# session, the encoder wins three of four - see the model cards.
+# Three backbones on the same head and the same task. The two 6M's are bidirectional BERT
+# encoders (one Arabic-only, one pretrained jointly on Arabic+English); the 52M is a Llama
+# decoder. Measured on the same eval sets in one session each - see the model cards.
 REGISTRY = {
-    "52M":     {"repo": "oddadmix/Nawah-Router-v3",         "label": "Nawah-Router-v3 (Llama decoder)"},
-    "6M-BERT": {"repo": "oddadmix/Nawah-Router-BERT-6M-v2", "label": "Nawah-Router-BERT-6M-v2 (BERT encoder)"},
+    "52M":         {"repo": "oddadmix/Nawah-Router-v3",
+                     "label": "Nawah-Router-v3 (Llama decoder, Arabic)"},
+    "6M-BERT":     {"repo": "oddadmix/Nawah-Router-BERT-6M-v2",
+                     "label": "Nawah-Router-BERT-6M-v2 (BERT encoder, Arabic)"},
+    "6M-BILINGUAL": {"repo": "oddadmix/Nawah-Router-BERT-6M-bilingual-pretrained",
+                      "label": "Nawah-Router-BERT-6M-bilingual-pretrained (BERT encoder, English+Arabic)"},
 }
 DEFAULT = os.environ.get("MODEL_KEY", "6M-BERT")
 
@@ -79,7 +83,7 @@ def api_route(req: RouteReq):
         # returned in the caller's order so the UI does not reshuffle rows under the cursor
         "results": [{"route": c, "score": order.get(c, 0.0)} for c in cats],
         "top": max(range(len(cats)), key=lambda i: order.get(cats[i], 0.0)),
-        "ms": round((time.perf_counter() - t0) * 1000), "tokens": ntok,
+        "ms": round((time.perf_counter() - t0) * 1000), "tokens": ntok, "lang": detect_lang(text),
         "model": sel["label"], "params": sel["params"]})