Update app.py

app.py

@@ -11,6 +11,10 @@ MODELS = {
         "repo": "olaverse/lid-neural-5",
         "note": "Yoruba, Hausa, Igbo, Nigerian Pidgin",
     },
+    "lid-neural-5.1 (Nigerian, 4 langs, sentence-level)": {
+        "repo": "olaverse/lid-neural-5.1",
+        "note": "Hausa, Yoruba, Igbo, Nigerian Pidgin — built on mist-encoder-base-ng, tuned for short/sentence-level text (97.6% acc)",
+    },
     "lid-lite-25 (fastText, 25 langs)": {
         "repo": "olaverse/lid-lite-25",
         "note": "CPU-only, sub-ms inference — trained on both long passages and short queries",
@@ -169,7 +173,7 @@ with gr.Blocks() as demo:
 
     with gr.Tab("Compare all models"):
         gr.Markdown(
-            "Runs the same text through all four models and shows top-3 predictions from each. "
+            "Runs the same text through all five models and shows top-3 predictions from each. "
             "Only one model is kept in memory at a time, so this reloads each checkpoint in turn — "
             "expect it to take a few seconds longer than the single-model tab."
         )
@@ -188,6 +192,7 @@ with gr.Blocks() as demo:
     ---
     **Model notes:**
     - `lid-neural-5` — Nigerian-focused, 4 languages (Yoruba, Hausa, Igbo, Pidgin)
+    - `lid-neural-5.1` — Nigerian-focused, same 4 languages, sentence-level tuned on `mist-encoder-base-ng` (97.6% acc; most residual error involves Pidgin, which shares vocabulary with the others)
     - `lid-lite-25` — fastText, CPU-only, 25 languages
     - `lid-neural-25.1` / `.2` — XLM-R fine-tunes, 25 languages, tuned for passages vs. short queries respectively.
       Known limitation across both: Zulu/Xhosa confusion on short text (see model cards).