Add Benchmarks tab with public dataset results

README.md

@@ -1,5 +1,5 @@
 ---
-title: GnaniNet Fraud Classifier
+title: KestrelNet Fraud Classifier
 emoji: 🔍
 colorFrom: yellow
 colorTo: red

app.py

@@ -1,5 +1,5 @@
 """
-GnaniNet Fraud Classifier — Gradio Space
+KestrelNet Fraud Classifier — Gradio Space
 
 Interactive demo for a 1,059-parameter fraud detection model.
 Pure NumPy inference, no GPU required.
@@ -21,7 +21,7 @@ def _softmax(logits):
     return e / e.sum()
 
 
-class GnaniNet:
+class KestrelNet:
     def __init__(self, input_dim, hidden_dims, output_dim):
         self.input_dim = input_dim
         self.hidden_dims = list(hidden_dims)
@@ -58,7 +58,7 @@ class GnaniNet:
 
 # ── Load model ──────────────────────────────────────────────────────────────
 
-model = GnaniNet(14, [32, 16], 3).load("weights.txt")
+model = KestrelNet(14, [32, 16], 3).load("weights.txt")
 
 
 # ── Feature normalization ───────────────────────────────────────────────────
@@ -190,8 +190,126 @@ A fully-connected classifier smaller than most models' bias vectors.
 </div>
 """
 
+BENCHMARKS_HTML = """
+<div style="max-width: 780px; margin: 0 auto; font-family: Inter, sans-serif;">
+
+<h3 style="margin-bottom: 4px;">Verified on Public Kaggle Datasets</h3>
+<p style="color: #888; font-size: 0.85em; margin-top: 0;">
+All datasets publicly available. Results independently reproducible. Pure NumPy, CPU only.
+</p>
+
+<table style="width:100%; border-collapse: collapse; font-size: 0.9em; margin: 16px 0;">
+<thead>
+<tr style="border-bottom: 2px solid #444; text-align: left;">
+    <th style="padding: 8px;">Dataset</th>
+    <th style="padding: 8px;">Task</th>
+    <th style="padding: 8px;">Accuracy</th>
+    <th style="padding: 8px;">F1 / AUC</th>
+    <th style="padding: 8px;">Params</th>
+    <th style="padding: 8px;">Latency</th>
+</tr>
+</thead>
+<tbody>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;"><strong>ECG Heartbeat</strong><br><span style="color:#888;font-size:0.8em;">MIT-BIH Arrhythmia</span></td>
+    <td style="padding: 8px;">5-class arrhythmia</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">97.2%</td>
+    <td style="padding: 8px;">F1 0.853</td>
+    <td style="padding: 8px;">12,756</td>
+    <td style="padding: 8px;">56 μs</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;"><strong>EEG Emotions</strong><br><span style="color:#888;font-size:0.8em;">Brainwave Sentiment</span></td>
+    <td style="padding: 8px;">3-class emotion</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">99.1%</td>
+    <td style="padding: 8px;">F1 0.991</td>
+    <td style="padding: 8px;">163,788</td>
+    <td style="padding: 8px;">1.3 ms</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;"><strong>EEG Eye State</strong><br><span style="color:#888;font-size:0.8em;">Roesler / UCI</span></td>
+    <td style="padding: 8px;">Binary open/closed</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">94.2%</td>
+    <td style="padding: 8px;">AUC 0.986</td>
+    <td style="padding: 8px;">1,576</td>
+    <td style="padding: 8px;">17 μs</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;"><strong>Seizure Prediction</strong><br><span style="color:#888;font-size:0.8em;">Bonn University EEG</span></td>
+    <td style="padding: 8px;">Binary seizure</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">97.1%</td>
+    <td style="padding: 8px;">AUC 0.988</td>
+    <td style="padding: 8px;">12,072</td>
+    <td style="padding: 8px;">—</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;"><strong>HAR Smartphones</strong><br><span style="color:#888;font-size:0.8em;">UCI Activity Recognition</span></td>
+    <td style="padding: 8px;">6-class activity</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">94.9%</td>
+    <td style="padding: 8px;">F1 0.949</td>
+    <td style="padding: 8px;">15,416</td>
+    <td style="padding: 8px;">70 μs</td>
+</tr>
+<tr style="border-bottom: 2px solid #444;">
+    <td style="padding: 8px;"><strong>Fraud Detection</strong><br><span style="color:#888;font-size:0.8em;">Proprietary</span></td>
+    <td style="padding: 8px;">3-class fraud</td>
+    <td style="padding: 8px; color: #22c55e; font-weight: 700;">91.6%</td>
+    <td style="padding: 8px;">—</td>
+    <td style="padding: 8px;">1,059</td>
+    <td style="padding: 8px;">5 μs</td>
+</tr>
+</tbody>
+</table>
+
+<h3 style="margin-bottom: 4px;">Parameter Efficiency vs Typical Models</h3>
+
+<table style="width:100%; border-collapse: collapse; font-size: 0.9em; margin: 16px 0;">
+<thead>
+<tr style="border-bottom: 2px solid #444; text-align: left;">
+    <th style="padding: 8px;">Dataset</th>
+    <th style="padding: 8px;">Typical CNN/LSTM</th>
+    <th style="padding: 8px;">Ours</th>
+    <th style="padding: 8px;">Reduction</th>
+</tr>
+</thead>
+<tbody>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;">ECG Heartbeat</td>
+    <td style="padding: 8px; color: #888;">500K – 2M</td>
+    <td style="padding: 8px; font-weight: 700;">12,756</td>
+    <td style="padding: 8px; color: #f59e0b; font-weight: 700;">40–160x smaller</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;">EEG Emotions</td>
+    <td style="padding: 8px; color: #888;">1M+</td>
+    <td style="padding: 8px; font-weight: 700;">163,788</td>
+    <td style="padding: 8px; color: #f59e0b; font-weight: 700;">6x smaller</td>
+</tr>
+<tr style="border-bottom: 1px solid #333;">
+    <td style="padding: 8px;">EEG Eye State</td>
+    <td style="padding: 8px; color: #888;">100K+</td>
+    <td style="padding: 8px; font-weight: 700;">1,576</td>
+    <td style="padding: 8px; color: #f59e0b; font-weight: 700;">63x smaller</td>
+</tr>
+<tr style="border-bottom: 2px solid #444;">
+    <td style="padding: 8px;">HAR Smartphones</td>
+    <td style="padding: 8px; color: #888;">200K – 1M</td>
+    <td style="padding: 8px; font-weight: 700;">15,416</td>
+    <td style="padding: 8px; color: #f59e0b; font-weight: 700;">13–65x smaller</td>
+</tr>
+</tbody>
+</table>
+
+<p style="color: #666; font-size: 0.8em; text-align: center; margin-top: 24px;">
+Two model families: <strong>KestrelNet</strong> (standard FC, minimal params) and <strong>GoshawkNet</strong> (multivector products, deeper pattern capture).<br>
+Named after raptors — bird size matches model size, hunting style matches classification style.
+</p>
+
+</div>
+"""
+
 with gr.Blocks(
-    title="GnaniNet Fraud Classifier",
+    title="KestrelNet Fraud Classifier",
     theme=gr.themes.Base(
         primary_hue=gr.themes.colors.orange,
         neutral_hue=gr.themes.colors.gray,
@@ -202,81 +320,86 @@ with gr.Blocks(
     footer { display: none !important; }
     """,
 ) as demo:
-    gr.Markdown("# GnaniNet", elem_id="title")
+    gr.Markdown("# KestrelNet", elem_id="title")
     gr.HTML(DESCRIPTION)
 
-    with gr.Row():
-        preset = gr.Dropdown(
-            choices=list(PRESETS.keys()),
-            label="Load preset",
-            interactive=True,
-            scale=2,
-        )
-        threshold = gr.Radio(
-            choices=["Standard", "Conservative", "Strict"],
-            value="Standard",
-            label="Threshold mode",
-            scale=3,
-        )
-
-    with gr.Row():
-        with gr.Column(scale=3):
-            gr.Markdown("### Transaction Features")
-            with gr.Row():
-                amount_ratio = gr.Number(label="Amount / 90-day avg", value=1.0, minimum=0, maximum=100)
-                hour = gr.Slider(label="Hour", value=14, minimum=0, maximum=23, step=1)
-                day_of_week = gr.Slider(label="Day of week (0=Mon)", value=2, minimum=0, maximum=6, step=1)
-
-            with gr.Row():
-                location_delta = gr.Number(label="Location delta (σ)", value=0.1, minimum=0, maximum=10)
-                velocity_1h = gr.Slider(label="Txns past hour", value=1, minimum=0, maximum=99, step=1)
-                velocity_24h = gr.Slider(label="Txns past 24h", value=3, minimum=0, maximum=999, step=1)
-
+    with gr.Tabs():
+        with gr.TabItem("Classify"):
             with gr.Row():
-                merchant_risk = gr.Slider(label="Merchant risk", value=0.05, minimum=0, maximum=1, step=0.01)
-                account_age_days = gr.Number(label="Account age (days)", value=1200, minimum=0, maximum=36500)
-                prev_fraud_score = gr.Slider(label="Prev fraud score", value=0.0, minimum=0, maximum=1, step=0.01)
+                preset = gr.Dropdown(
+                    choices=list(PRESETS.keys()),
+                    label="Load preset",
+                    interactive=True,
+                    scale=2,
+                )
+                threshold = gr.Radio(
+                    choices=["Standard", "Conservative", "Strict"],
+                    value="Standard",
+                    label="Threshold mode",
+                    scale=3,
+                )
 
             with gr.Row():
-                international = gr.Checkbox(label="International", value=False)
-                card_present = gr.Checkbox(label="Card present", value=True)
-                device_match = gr.Checkbox(label="Device match", value=True)
-
-        with gr.Column(scale=2):
-            gr.Markdown("### Result")
-            verdict_output = gr.HTML()
-            scores_output = gr.Label(label="Class probabilities", num_top_classes=3)
-
-    feature_inputs = [
-        amount_ratio, hour, day_of_week, location_delta,
-        velocity_1h, velocity_24h, merchant_risk,
-        international, card_present, device_match,
-        account_age_days, prev_fraud_score,
-    ]
-
-    all_inputs = feature_inputs + [threshold]
-
-    # Preset loader
-    preset.change(fn=apply_preset, inputs=[preset], outputs=feature_inputs)
-
-    # Auto-classify on any change
-    for inp in all_inputs:
-        inp.change(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
-
-    # Classify button as fallback
-    btn = gr.Button("Classify", variant="primary")
-    btn.click(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
-
-    # Run once on load
-    demo.load(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
+                with gr.Column(scale=3):
+                    gr.Markdown("### Transaction Features")
+                    with gr.Row():
+                        amount_ratio = gr.Number(label="Amount / 90-day avg", value=1.0, minimum=0, maximum=100)
+                        hour = gr.Slider(label="Hour", value=14, minimum=0, maximum=23, step=1)
+                        day_of_week = gr.Slider(label="Day of week (0=Mon)", value=2, minimum=0, maximum=6, step=1)
+
+                    with gr.Row():
+                        location_delta = gr.Number(label="Location delta (σ)", value=0.1, minimum=0, maximum=10)
+                        velocity_1h = gr.Slider(label="Txns past hour", value=1, minimum=0, maximum=99, step=1)
+                        velocity_24h = gr.Slider(label="Txns past 24h", value=3, minimum=0, maximum=999, step=1)
+
+                    with gr.Row():
+                        merchant_risk = gr.Slider(label="Merchant risk", value=0.05, minimum=0, maximum=1, step=0.01)
+                        account_age_days = gr.Number(label="Account age (days)", value=1200, minimum=0, maximum=36500)
+                        prev_fraud_score = gr.Slider(label="Prev fraud score", value=0.0, minimum=0, maximum=1, step=0.01)
+
+                    with gr.Row():
+                        international = gr.Checkbox(label="International", value=False)
+                        card_present = gr.Checkbox(label="Card present", value=True)
+                        device_match = gr.Checkbox(label="Device match", value=True)
+
+                with gr.Column(scale=2):
+                    gr.Markdown("### Result")
+                    verdict_output = gr.HTML()
+                    scores_output = gr.Label(label="Class probabilities", num_top_classes=3)
+
+            feature_inputs = [
+                amount_ratio, hour, day_of_week, location_delta,
+                velocity_1h, velocity_24h, merchant_risk,
+                international, card_present, device_match,
+                account_age_days, prev_fraud_score,
+            ]
+
+            all_inputs = feature_inputs + [threshold]
+
+            # Preset loader
+            preset.change(fn=apply_preset, inputs=[preset], outputs=feature_inputs)
+
+            # Auto-classify on any change
+            for inp in all_inputs:
+                inp.change(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
+
+            # Classify button as fallback
+            btn = gr.Button("Classify", variant="primary")
+            btn.click(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
+
+            # Run once on load
+            demo.load(fn=classify, inputs=all_inputs, outputs=[verdict_output, scores_output])
+
+        with gr.TabItem("Benchmarks"):
+            gr.HTML(BENCHMARKS_HTML)
 
     gr.Markdown("""
 ---
 <div style="text-align:center; font-size: 0.8em; color: #888;">
 Architecture: 14 → 32 → 16 → 3 &nbsp;·&nbsp; ReLU activations &nbsp;·&nbsp; Softmax output &nbsp;·&nbsp; Analytic backprop training<br>
 No PyTorch. No TensorFlow. No ONNX. Just NumPy.<br><br>
-<a href="https://huggingface.co/gnaninet/fraud-classifier">Model Card</a> &nbsp;·&nbsp;
-<a href="https://gnaninet.ai">Website</a>
+<a href="https://huggingface.co/reddysama/gnaninet-fraud-classifier">Model Card</a> &nbsp;·&nbsp;
+<a href="https://naninet.ai">Website</a>
 </div>
     """)