feat: implement category-specific benchmark scores and fix deployment

Author: voidful

.gitignore

@@ -31,4 +31,5 @@ benchmarking.py
 
 # Data
 cached_datasets/
-datasets/
+datasets/
+*_evaluation_results*.json

SoundCodec/dataset/__init__.py

@@ -8,12 +8,12 @@ def load_dataset(dataset_name):
         # Fallback to loading from Hugging Face Hub
         ds = hf_load_dataset(dataset_name)
         if isinstance(ds, dict):
-            if "test" in ds:
-                return ds["test"]
-            if "validation" in ds:
-                return ds["validation"]
-            if "train" in ds:
-                return ds["train"]
-            # return the first split if none of the above are found
-            return ds[list(ds.keys())[0]]
+            from datasets import concatenate_datasets
+            all_ds = []
+            for split_name, split_ds in ds.items():
+                # Add category column if it doesn't exist
+                if 'category' not in split_ds.column_names:
+                    split_ds = split_ds.add_column('category', [split_name] * len(split_ds))
+                all_ds.append(split_ds)
+            return concatenate_datasets(all_ds).shuffle(seed=42)
         return ds

update_leaderboard.py

@@ -1,7 +1,15 @@
 import json
 
-# Load benchmark results
-with open('._datasets_voidful_codec-superb-tiny_synth_evaluation_results_20251218_204458.json', 'r') as f:
+import glob
+import os
+
+# Load latest benchmark results
+json_files = glob.glob('*codec-superb-tiny_synth_evaluation_results*.json')
+if not json_files:
+    raise FileNotFoundError("No benchmark results found.")
+latest_file = max(json_files, key=os.path.getmtime)
+print(f"Loading results from {latest_file}")
+with open(latest_file, 'r') as f:
     benchmark_results = json.load(f)
 
 # Hardcoded BPS mapping (bitrate in kbps or as used in data.js)
@@ -47,16 +55,32 @@
 
 new_results = {}
 
-for model_name, metrics in benchmark_results.items():
+new_results = {}
+
+for model_name, metrics_data in benchmark_results.items():
     entry = {
         'bps': bps_mapping.get(model_name, 0)
     }
-    for m in metrics_to_include:
-        val = metrics.get(m, 0)
-        # Handle NaN
-        if val != val: # NaN check
-            val = 0
-        entry[m] = round(float(val), 3)
+    
+    # Check if nested by category
+    is_nested = any(isinstance(v, dict) for v in metrics_data.values())
+    
+    if is_nested:
+        for category, metrics in metrics_data.items():
+            for m in metrics_to_include:
+                val = metrics.get(m, 0)
+                if val != val: # NaN check
+                    val = 0
+                entry[f"{category.lower()}_{m}"] = round(float(val), 3)
+    else:
+        # Legacy format: metrics_data is {metric: value}
+        # Map to 'overall' category by default
+        for m in metrics_to_include:
+            val = metrics_data.get(m, 0)
+            if val != val: # NaN check
+                val = 0
+            entry[f"overall_{m}"] = round(float(val), 3)
+            
     new_results[model_name] = entry
 
 # Format as JavaScript

web/src/App.css

@@ -15,6 +15,17 @@
   z-index: 10;
 }
 
+.leaderboard-note {
+  margin-bottom: 1.5rem;
+  color: var(--text-secondary);
+  font-size: 0.9375rem;
+}
+
+.leaderboard-note strong {
+  color: var(--accent-color);
+  font-weight: 600;
+}
+
 .main-footer {
   padding: 3rem;
   text-align: center;

web/src/App.js

@@ -31,6 +31,7 @@ function App() {
           </Card>
 
           <Card title="Leaderboard" delay={0.4}>
+            <p className="leaderboard-note">Comparing performance across <strong>Speech</strong>, <strong>Audio</strong>, and <strong>Music</strong> categories.</p>
             <div className="results-section">
               <Leaderboard results={results} />
             </div>

web/src/Leaderboard.css

@@ -20,11 +20,23 @@ th {
   font-weight: 600;
   color: var(--text-secondary);
   border-bottom: 1px solid var(--surface-border);
+  border-right: 1px solid rgba(255, 255, 255, 0.05);
   cursor: pointer;
   transition: var(--transition-fast);
   white-space: nowrap;
 }
 
+thead tr:first-child th {
+  border-bottom: 1px solid rgba(255, 255, 255, 0.1);
+  background: rgba(255, 255, 255, 0.05);
+  color: var(--text-primary);
+  text-align: center;
+  font-size: 0.75rem;
+  letter-spacing: 0.1em;
+  text-transform: uppercase;
+  padding: 0.5rem 1rem;
+}
+
 th:hover {
   background: rgba(255, 255, 255, 0.06);
   color: var(--text-primary);

web/src/Leaderboard.js

@@ -16,17 +16,41 @@ const Leaderboard = ({ results }) => {
 
   const columns = React.useMemo(() => {
     const firstItem = results[Object.keys(results)[0]];
-    return [
+    const categories = ['Speech', 'Audio', 'Music', 'Overall'];
+    const metrics_keys = ['mel', 'pesq', 'stoi', 'f0corr'];
+
+    const colGroups = [
       {
-        Header: 'Model',
-        accessor: 'model',
-      },
-      ...Object.keys(firstItem).map(key => ({
-        Header: key.toUpperCase(),
-        accessor: key,
-        sortType: 'basic',
-      })),
+        Header: 'Model Info',
+        columns: [
+          { Header: 'Model', accessor: 'model' },
+          { Header: 'BPS', accessor: 'bps' }
+        ]
+      }
     ];
+
+    categories.forEach(cat => {
+      const catColumns = metrics_keys.map(m => {
+        const key = `${cat.toLowerCase()}_${m}`;
+        if (key in firstItem || true) { // Force inclusion or check existence
+          return {
+            Header: m.toUpperCase(),
+            accessor: key,
+            sortType: 'basic',
+          };
+        }
+        return null;
+      }).filter(Boolean);
+
+      if (catColumns.length > 0) {
+        colGroups.push({
+          Header: cat.toUpperCase(),
+          columns: catColumns
+        });
+      }
+    });
+
+    return colGroups;
   }, [results]);
 
   const {