load version thresholds

app.py

@@ -34,6 +34,22 @@ def log_odds(p, eps=0.0):
     return np.log(p / (1 - p + eps))
 
 
+def get_threshold(rev_id):
+    import requests    
+
+    DEFAULT_THRESH = 0.0049
+    URL = f"https://huggingface.co/livekit/turn-detector/resolve/{rev_id}/languages.json"
+
+    try:
+        config = requests.get(URL).json().get("en")
+    except Exception as e:
+        print(f"Error loading languages.json: \n{e}")
+        config = {}
+    
+    return config.get("threshold", DEFAULT_THRESH)
+
+
+
 # ===== Per-model runner (keeps tokenizer/model and token ids) =====
 class ModelRunner:
     def __init__(self, model_id: str, revision: str | None = None, dtype=torch.bfloat16):
@@ -48,6 +64,7 @@ class ModelRunner:
             device_map="auto",
         )
         self.model.eval()
+        self.thresh = get_thresh(revision)
 
         # Pull commonly used tokens, falling back gracefully if not present
         self.START_TOKEN_ID = self._tok_id("<|im_start|>")
@@ -149,6 +166,8 @@ class ModelRunner:
 
     def make_styled_df(self, df: pd.DataFrame, thresh: float, cmap="coolwarm") -> str:
         EPS = 1e-12
+        thresh = self.thresh
+        
         _df = df.copy()
         _df = _df[~_df.token.isin(self.CONTROL_TOKS)]
         _df.token = _df.token.replace({"\n": "⏎", " ": "␠"})
@@ -185,6 +204,7 @@ class ModelRunner:
     def generate_highlighted_text(self, text: str, threshold: float):
         """Returns: (highlighted_list, styled_html) for Gradio"""
         eps = 1e-12
+        threshold = self.thresh
         if not text:
             return [], "<div>No input.</div>"
 
@@ -213,7 +233,6 @@ def get_runner(model_id: str, revision: str | None):
 
 
 # ===== Gradio App =====
-EN_THRESHOLD = 0.0049
 
 import spaces
 
@@ -221,17 +240,15 @@ import spaces
 def compare_models(
     text: str,
     model_a_id: str,
-    model_a_rev: str,
-    thresh_a: float,
+    model_a_rev: str,    
     model_b_id: str,
-    model_b_rev: str,
-    thresh_b: float,
+    model_b_rev: str,    
 ):
     runner_a = get_runner(model_a_id, model_a_rev if model_a_rev else None)
     runner_b = get_runner(model_b_id, model_b_rev if model_b_rev else None)
 
-    ht_a, html_a = runner_a.generate_highlighted_text(text, threshold=thresh_a)
-    ht_b, html_b = runner_b.generate_highlighted_text(text, threshold=thresh_b)
+    ht_a, html_a = runner_a.generate_highlighted_text(text)
+    ht_b, html_b = runner_b.generate_highlighted_text(text)
 
     # Optional: prepend small headers indicating model names in the HTML blocks
     html_a = f"<h4 style='margin:0 0 8px 0'>{model_a_id}@{model_a_rev or 'default'}</h4>" + html_a
@@ -265,13 +282,12 @@ Red ⇒ agent should reply • Blue ⇒ agent should wait"""
             gr.Markdown("### Model A")
             model_a_id = gr.Textbox(value=DEFAULT_MODEL_A_ID, label="Model ID")
             model_a_rev = gr.Textbox(value=DEFAULT_MODEL_A_REV, label="Revision (optional)")
-            thresh_a = gr.Slider(0.0001, 0.05, value=EN_THRESHOLD, step=0.0001, label="Threshold")
+            
 
         with gr.Column():
             gr.Markdown("### Model B")
             model_b_id = gr.Textbox(value=DEFAULT_MODEL_B_ID, label="Model ID")
-            model_b_rev = gr.Textbox(value=DEFAULT_MODEL_B_REV, label="Revision (optional)")
-            thresh_b = gr.Slider(0.0001, 0.05, value=EN_THRESHOLD, step=0.0001, label="Threshold")
+            model_b_rev = gr.Textbox(value=DEFAULT_MODEL_B_REV, label="Revision (optional)")            
 
     run_btn = gr.Button("Run Comparison", variant="primary")
 
@@ -293,7 +309,7 @@ Red ⇒ agent should reply • Blue ⇒ agent should wait"""
 
     run_btn.click(
         fn=compare_models,
-        inputs=[text_in, model_a_id, model_a_rev, thresh_a, model_b_id, model_b_rev, thresh_b],
+        inputs=[text_in, model_a_id, model_a_rev, model_b_id, model_b_rev],
         outputs=[out_ht_a, out_html_a, out_ht_b, out_html_b]
     )