Upload index.html with huggingface_hub

index.html

@@ -194,6 +194,28 @@
         }
         .data-table .num { font-family: 'JetBrains Mono', monospace; font-size: 0.82rem; }
         .data-table .muted { color: var(--text-muted); }
+        .footnote {
+            color: var(--text-muted); font-size: 0.82rem;
+            margin: 0.5rem 0 1.5rem 0; line-height: 1.5;
+        }
+        .footnote a { color: var(--text-muted); }
+        .drill-down {
+            background: var(--bg-card);
+            border-left: 4px solid var(--gold);
+            border-radius: 6px;
+            padding: 1rem 1.2rem;
+            margin: 1.2rem 0;
+        }
+        .drill-down p { margin: 0.3rem 0; font-size: 0.92rem; }
+        .drill-down .cta-link {
+            display: inline-block;
+            margin-top: 0.4rem;
+            font-weight: 600;
+            color: var(--navy);
+        }
+        @media (prefers-color-scheme: dark) {
+            .drill-down .cta-link { color: var(--gold); }
+        }
         /* Model cards */
         .model-grid {
             display: grid;
@@ -296,23 +318,56 @@
             </ul>
         </nav>
 
-        <!-- Overview: All families comparison -->
-        <h2>Detection &mdash; Nano Accuracy Comparison</h2>
-        <p>ONNX FP32 mAP@0.5 on COCO val2017 (5000 images, 80 classes). Nano size for each family.</p>
-        <div class="bar-chart">
-            <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:91.5%"><span>54.9%</span></div></div></div>
-            <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:89.0%"><span>53.4%</span></div></div></div>
-            <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:83.6%"><span>50.2%</span></div></div></div>
-            <div class="bar-row"><span class="bar-label">YOLOv5</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:82.6%"><span>49.6%</span></div></div></div>
+        <!-- ═══════════════════════════════════════════════════════ -->
+        <!-- EXECUTIVE SUMMARY: VALIDATION PIPELINE + PLATFORM       -->
+        <!-- ═══════════════════════════════════════════════════════ -->
+
+        <h2>Validation Pipeline</h2>
+        <div class="diagram-container">
+            <img src="01-ecosystem.png" alt="EdgeFirst Model Zoo Ecosystem">
         </div>
+        <p>
+            Every artifact in the Model Zoo is measured on the <strong>same dataset on the same hardware</strong> users deploy on. <a href="https://edgefirst.studio"><strong>EdgeFirst Studio</strong></a> manages datasets, training, multi-format export, and reference validation; on-target runs happen on a board farm of i.MX 8M Plus, i.MX 95, Ara-240, Hailo, and Jetson devices. Accuracy numbers and per-stage timing are pushed back to Studio session metrics and consumed by this Model Zoo when generating each model card.
+        </p>
 
-        <h2>Segmentation &mdash; Nano Accuracy Comparison</h2>
-        <p>ONNX FP32 Mask mAP@0.5-0.95 on COCO val2017. Nano size, split-decoder.</p>
-        <div class="bar-chart">
-            <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:92.5%"><span>37.0%</span></div></div></div>
-            <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:88.7%"><span>35.5%</span></div></div></div>
-            <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:85.2%"><span>34.1%</span></div></div></div>
+        <h3>End-to-end Flow</h3>
+        <div class="diagram-container">
+            <img src="02-model-lifecycle.png" alt="Model Lifecycle: 5 stages from training to publication">
         </div>
+        <p>
+            Each training session produces a single set of weights. The export pipeline emits ONNX FP32, INT8 TFLite, and platform-specific compiled formats (i.MX 95 Neutron, NXP Ara-240 DVM, Hailo HEF, Jetson TensorRT). Every output is paired with an on-target validation run that captures both <strong>accuracy</strong> (COCO/LVIS mAP against the validation set) and <strong>full-pipeline timing</strong>. The ONNX FP32 run from each training session serves as the reference baseline; quantization and runtime loss are measured relative to it, not relative to externally-published numbers.
+        </p>
+
+        <h3>EdgeFirst Profiler</h3>
+        <div class="diagram-container">
+            <img src="03-on-target-validation.png" alt="On-Target Validation Pipeline">
+        </div>
+        <p>
+            The <strong>EdgeFirst Profiler</strong> is the on-target agent that drives every validation. Given a model and a dataset, it runs full inference on the target device, captures per-image predictions in the EdgeFirst Arrow/Parquet format, and emits a <a href="https://perfetto.dev"><strong>Perfetto</strong></a> trace alongside the predictions. The Profiler is hardware-aware: it loads each runtime through its native delegate &mdash; Verisilicon VX Delegate on i.MX 8M Plus, eIQ Neutron Delegate on i.MX 95, Kinara SDK on Ara-240, HailoRT on RPi5 + Hailo, TensorRT on Jetson &mdash; so every stage timed by the trace corresponds to what a deployed application would experience.
+        </p>
+
+        <h3>EdgeFirst Validator</h3>
+        <p>
+            The <strong>EdgeFirst Validator</strong> is the off-target post-processor. It consumes the Profiler's predictions and Perfetto trace, computes the full 12-metric COCO accuracy tuple via <code>pycocotools</code> (or <code>lvis-api</code> for large-vocabulary datasets), and rebuilds per-stage timing summaries from the trace. Results land in a structured YAML payload attached to the Studio validation session &mdash; the same payload the Model Zoo reads to render this page. Accuracy and timing are computed independently of the runtime that produced the predictions, so toolchain regressions surface as cross-platform divergence rather than silent failures.
+        </p>
+
+        <h3>EdgeFirst HAL</h3>
+        <p>
+            The <a href="https://github.com/EdgeFirstAI/hal"><strong>EdgeFirst Hardware Abstraction Layer</strong> (HAL)</a> provides the hardware-accelerated primitives used at both validation and deployment time. The Profiler uses HAL for letterbox resize, color-space conversion, normalization, layout conversion, and post-decode (YOLO/ModelPack output decoding, NMS, mask materialisation). HAL automatically selects DMA-BUF, OpenGL ES, NXP G2D, or CPU paths depending on the platform &mdash; so the timing measured during validation reflects the same accelerated path a production runtime would take. HAL ships as a Rust crate, a Python package, and a C library under Apache 2.0.
+        </p>
+
+        <h3>Latency and Pipelined Throughput</h3>
+        <p>
+            Two complementary timing surfaces are reported per validation:
+        </p>
+        <table class="data-table">
+            <tr><th>Surface</th><th>What it captures</th><th>When it&#39;s present</th></tr>
+            <tr><td><code>timing.inline</code></td><td>Per-image <code>preprocess_ms</code>, <code>inference_ms</code>, <code>postprocess_ms</code> with min / mean / median / p95 / p99 / max</td><td>Always &mdash; the universal contract every producer fills in</td></tr>
+            <tr><td><code>timing.trace</code></td><td>Full per-stage breakdown from the Perfetto trace (typically 25&ndash;33 stages including delegate work, tensor moves, decode passes, NMS, parquet flush), plus end-to-end FPS distribution</td><td>When the Profiler emits a sidecar trace (almost all runs)</td></tr>
+        </table>
+        <p>
+            <strong>Throughput exceeds the sum of stage latencies</strong> because the runtime pipelines I/O, host preprocessing, NPU inference, and decode across frames. Reporting <code>1000 / (preprocess + inference + postprocess)</code> understates real throughput; the Model Zoo uses the <strong>measured end-to-end FPS</strong> from the trace (<code>trace.fps.median</code>) as the headline number reported in every per-target table on every model card. As a concrete example, YOLOv5n on i.MX 95 Neutron has per-stage means 21.7 ms preprocess + 12.2 ms inference + 15.8 ms postprocess (naive estimate ~20 FPS), but the measured pipelined throughput is <strong>56 FPS median</strong> &mdash; the 2.8&times; gap is the value the pipelining delivers.
+        </p>
 
         <h2>Platform Support</h2>
         <div class="badges">
@@ -324,13 +379,35 @@
         </div>
 
         <table class="data-table">
-            <tr><th>Family</th><th>ONNX</th><th>TFLite</th><th>i.MX 93</th><th>i.MX 95</th><th>Ara240</th><th>Hailo</th><th>Jetson</th></tr>
-            <tr><td>YOLO26</td><td>✓</td><td>✓</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            <tr><td>YOLO11</td><td>✓</td><td>✓</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            <tr><td>YOLOv8</td><td>✓</td><td>✓</td><td class="muted">&mdash;</td><td><span class="wip-tag">WIP</span></td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            <tr><td>YOLOv5</td><td>✓</td><td>✓</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
+            <tr><th>Family</th><th>ONNX</th><th>i.MX 8M Plus</th><th>i.MX 95</th><th>Ara240</th><th>Hailo</th><th>Jetson</th></tr>
+            <tr><td>YOLO26</td><td>✓</td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓</td><td>✓</td><td>✓</td></tr>
+            <tr><td>YOLO11</td><td>✓</td><td>✓</td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓</td><td>✓</td><td>✓</td></tr>
+            <tr><td>YOLOv8</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr>
+            <tr><td>YOLOv5</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr>
         </table>
 
+        <p id="platform-notes" class="footnote">
+            ✓ indicates a compiled model artifact is published in the corresponding repo. <sup>[1]</sup> Some sizes for this family / platform combination currently render an em-dash in the model card's <em>On-target validation results</em> table — see the per-repo card for the linked Studio validation session (<code>v-XXXX</code>) and its current status.
+        </p>
+
+        <!-- Overview: All families comparison -->
+        <h2>Detection &mdash; Nano Accuracy</h2>
+        <p>ONNX FP32 mAP@0.5 on COCO val2017 (5000 images, 80 classes). Nano size for each family. Source: EdgeFirst Studio validation sessions cited in each model card.</p>
+        <div class="bar-chart">
+            <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:91.8%"><span>55.06%</span></div></div></div>
+            <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:88.4%"><span>53.05%</span></div></div></div>
+            <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:84.3%"><span>50.55%</span></div></div></div>
+            <div class="bar-row"><span class="bar-label">YOLOv5</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:79.6%"><span>47.77%</span></div></div></div>
+        </div>
+
+        <h2>Segmentation &mdash; Nano Mask Accuracy</h2>
+        <p>ONNX FP32 Mask mAP@0.5-0.95 on COCO val2017. Nano size, split-decoder. Source: EdgeFirst Studio validation sessions cited in each model card.</p>
+        <div class="bar-chart">
+            <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:78.0%"><span>31.21%</span></div></div></div>
+            <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:71.9%"><span>28.75%</span></div></div></div>
+            <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:68.5%"><span>27.41%</span></div></div></div>
+        </div>
+
         <!-- ═══════════════════════════════════════════════════════ -->
         <!-- LEVEL 2: MODEL FAMILIES                                -->
         <!-- ═══════════════════════════════════════════════════════ -->
@@ -345,44 +422,46 @@
             <div class="model-grid">
                 <div class="model-card">
                     <h3><a href="#yolo26-det" target="_self">Detection</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano mAP@0.5: 54.9% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-det">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 55.06% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-det">HF Repo</a></p>
                 </div>
                 <div class="model-card">
                     <h3><a href="#yolo26-seg" target="_self">Segmentation</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano Mask mAP: 37.0% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-seg">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 31.21% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-seg">HF Repo</a></p>
                 </div>
             </div>
 
-            <h4>Size Scaling &mdash; Detection (ONNX FP32)</h4>
-            <table class="data-table">
-                <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th></tr>
-                <tr><td>Nano</td><td class="num">2.7M</td><td class="num">7.6</td><td class="num">54.9%</td><td class="num">39.7%</td></tr>
-                <tr><td>Small</td><td class="num">10.3M</td><td class="num">27.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Medium</td><td class="num">24.5M</td><td class="num">74.4</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Large</td><td class="num">42.5M</td><td class="num">155.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>XLarge</td><td class="num">67.5M</td><td class="num">244.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            </table>
-
             <!-- Level 3: YOLO26 Detection -->
             <div id="yolo26-det">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo26" target="_self">YOLO26</a> &rsaquo; <strong>Detection</strong></p>
-                <h3>YOLO26 Detection</h3>
-                <p>Accuracy on COCO val2017 (5000 images, 80 classes). <a href="https://huggingface.co/EdgeFirst/yolo26-det">View on HuggingFace &rarr;</a></p>
+                <h3>YOLO26 Detection &mdash; ONNX FP32 reference</h3>
+                <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX FP32 mAP@0.5</th><th>TFLite INT8 mAP@0.5</th><th>INT8 Drop</th></tr>
-                    <tr><td>Nano</td><td class="num">54.9%</td><td class="num">51.5%</td><td class="num">−3.4 pp</td></tr>
+                    <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
+                    <tr><td>Nano</td><td class="num">2.7M</td><td class="num">7.6</td><td class="num">55.06%</td><td class="num">39.71%</td><td class="num">42.87%</td></tr>
+                    <tr><td>Small</td><td class="num">10.3M</td><td class="num">27.0</td><td class="num">63.60%</td><td class="num">47.16%</td><td class="num">51.14%</td></tr>
+                    <tr><td>Medium</td><td class="num">24.5M</td><td class="num">74.4</td><td class="num">68.89%</td><td class="num">51.88%</td><td class="num">56.41%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo26-det">View YOLO26 Detection model card on HuggingFace &rarr;</a>
+                </div>
             </div>
 
             <!-- Level 3: YOLO26 Segmentation -->
             <div id="yolo26-seg">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo26" target="_self">YOLO26</a> &rsaquo; <strong>Segmentation</strong></p>
-                <h3>YOLO26 Segmentation</h3>
-                <p>Split-decoder architecture. <a href="https://huggingface.co/EdgeFirst/yolo26-seg">View on HuggingFace &rarr;</a></p>
+                <h3>YOLO26 Segmentation &mdash; ONNX FP32 reference</h3>
+                <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX Det mAP</th><th>INT8 Det mAP</th><th>ONNX Mask mAP</th><th>INT8 Mask mAP</th></tr>
-                    <tr><td>Nano</td><td class="num">29.6%</td><td class="num">26.8%</td><td class="num">37.0%</td><td class="num">34.5%</td></tr>
+                    <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
+                    <tr><td>Nano</td><td class="num">54.85%</td><td class="num">39.31%</td><td class="num">50.77%</td><td class="num">31.21%</td></tr>
+                    <tr><td>Small</td><td class="num">63.15%</td><td class="num">46.66%</td><td class="num">58.33%</td><td class="num">36.07%</td></tr>
+                    <tr><td>Medium</td><td class="num">68.80%</td><td class="num">51.76%</td><td class="num">63.36%</td><td class="num">39.47%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo26-seg">View YOLO26 Segmentation model card on HuggingFace &rarr;</a>
+                </div>
             </div>
             <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
         </div>
@@ -391,48 +470,50 @@
         <div class="family-section" id="yolo11">
             <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLO11</strong></p>
             <h2>YOLO11</h2>
-            <p>Newer architecture with attention blocks. Strong balance of accuracy and efficiency.</p>
+            <p>Architecture with C3k2 attention blocks. Anchor-free DFL detection head.</p>
 
             <h4>Tasks</h4>
             <div class="model-grid">
                 <div class="model-card">
                     <h3><a href="#yolo11-det" target="_self">Detection</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano mAP@0.5: 53.4% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-det">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 53.05% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-det">HF Repo</a></p>
                 </div>
                 <div class="model-card">
                     <h3><a href="#yolo11-seg" target="_self">Segmentation</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano Mask mAP: 35.5% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-seg">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 28.75% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-seg">HF Repo</a></p>
                 </div>
             </div>
 
-            <h4>Size Scaling &mdash; Detection (ONNX FP32)</h4>
-            <table class="data-table">
-                <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th></tr>
-                <tr><td>Nano</td><td class="num">2.6M</td><td class="num">6.5</td><td class="num">53.4%</td><td class="num">37.9%</td></tr>
-                <tr><td>Small</td><td class="num">9.4M</td><td class="num">21.5</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Medium</td><td class="num">20.1M</td><td class="num">68.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Large</td><td class="num">25.3M</td><td class="num">87.6</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>XLarge</td><td class="num">56.9M</td><td class="num">195.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            </table>
-
             <div id="yolo11-det">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo11" target="_self">YOLO11</a> &rsaquo; <strong>Detection</strong></p>
-                <h3>YOLO11 Detection</h3>
-                <p><a href="https://huggingface.co/EdgeFirst/yolo11-det">View on HuggingFace &rarr;</a></p>
+                <h3>YOLO11 Detection &mdash; ONNX FP32 reference</h3>
+                <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX FP32 mAP@0.5</th><th>TFLite INT8 mAP@0.5</th><th>INT8 Drop</th></tr>
-                    <tr><td>Nano</td><td class="num">53.4%</td><td class="num">50.1%</td><td class="num">−3.3 pp</td></tr>
+                    <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
+                    <tr><td>Nano</td><td class="num">2.6M</td><td class="num">6.5</td><td class="num">53.05%</td><td class="num">37.83%</td><td class="num">41.07%</td></tr>
+                    <tr><td>Small</td><td class="num">9.4M</td><td class="num">21.5</td><td class="num">61.21%</td><td class="num">44.91%</td><td class="num">48.44%</td></tr>
+                    <tr><td>Medium</td><td class="num">20.1M</td><td class="num">68.0</td><td class="num">65.91%</td><td class="num">49.59%</td><td class="num">53.80%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo11-det">View YOLO11 Detection model card on HuggingFace &rarr;</a>
+                </div>
             </div>
 
             <div id="yolo11-seg">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo11" target="_self">YOLO11</a> &rsaquo; <strong>Segmentation</strong></p>
-                <h3>YOLO11 Segmentation</h3>
-                <p><a href="https://huggingface.co/EdgeFirst/yolo11-seg">View on HuggingFace &rarr;</a></p>
+                <h3>YOLO11 Segmentation &mdash; ONNX FP32 reference</h3>
+                <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX Det mAP</th><th>INT8 Det mAP</th><th>ONNX Mask mAP</th><th>INT8 Mask mAP</th></tr>
-                    <tr><td>Nano</td><td class="num">28.4%</td><td class="num">27.1%</td><td class="num">35.5%</td><td class="num">34.4%</td></tr>
+                    <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
+                    <tr><td>Nano</td><td class="num">52.00%</td><td class="num">37.22%</td><td class="num">47.88%</td><td class="num">28.75%</td></tr>
+                    <tr><td>Small</td><td class="num">60.60%</td><td class="num">44.56%</td><td class="num">55.46%</td><td class="num">33.33%</td></tr>
+                    <tr><td>Medium</td><td class="num">65.91%</td><td class="num">49.59%</td><td class="num">60.54%</td><td class="num">36.61%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo11-seg">View YOLO11 Segmentation model card on HuggingFace &rarr;</a>
+                </div>
             </div>
             <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
         </div>
@@ -441,48 +522,50 @@
         <div class="family-section" id="yolov8">
             <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLOv8</strong></p>
             <h2>YOLOv8</h2>
-            <p>Anchor-free DFL detection head. Available in detection and instance-segmentation variants.</p>
+            <p>Anchor-free DFL detection head. Detection and instance-segmentation variants.</p>
 
             <h4>Tasks</h4>
             <div class="model-grid">
                 <div class="model-card">
                     <h3><a href="#yolov8-det" target="_self">Detection</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano mAP@0.5: 50.2% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-det">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 50.55% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-det">HF Repo</a></p>
                 </div>
                 <div class="model-card">
                     <h3><a href="#yolov8-seg" target="_self">Segmentation</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano Mask mAP: 34.1% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-seg">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 27.41% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-seg">HF Repo</a></p>
                 </div>
             </div>
 
-            <h4>Size Scaling &mdash; Detection (ONNX FP32)</h4>
-            <table class="data-table">
-                <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th></tr>
-                <tr><td>Nano</td><td class="num">3.2M</td><td class="num">8.9</td><td class="num">50.2%</td><td class="num">35.8%</td></tr>
-                <tr><td>Small</td><td class="num">11.2M</td><td class="num">28.8</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Medium</td><td class="num">25.9M</td><td class="num">79.3</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Large</td><td class="num">43.7M</td><td class="num">165.7</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>XLarge</td><td class="num">68.2M</td><td class="num">258.5</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            </table>
-
             <div id="yolov8-det">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov8" target="_self">YOLOv8</a> &rsaquo; <strong>Detection</strong></p>
-                <h3>YOLOv8 Detection</h3>
-                <p><a href="https://huggingface.co/EdgeFirst/yolov8-det">View on HuggingFace &rarr;</a></p>
+                <h3>YOLOv8 Detection &mdash; ONNX FP32 reference</h3>
+                <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX FP32 mAP@0.5</th><th>TFLite INT8 mAP@0.5</th><th>INT8 Drop</th></tr>
-                    <tr><td>Nano</td><td class="num">50.2%</td><td class="num">47.3%</td><td class="num">−2.9 pp</td></tr>
+                    <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
+                    <tr><td>Nano</td><td class="num">3.2M</td><td class="num">8.9</td><td class="num">50.55%</td><td class="num">35.87%</td><td class="num">38.92%</td></tr>
+                    <tr><td>Small</td><td class="num">11.2M</td><td class="num">28.8</td><td class="num">59.39%</td><td class="num">43.23%</td><td class="num">46.69%</td></tr>
+                    <tr><td>Medium</td><td class="num">25.9M</td><td class="num">79.3</td><td class="num">64.70%</td><td class="num">48.46%</td><td class="num">52.62%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov8-det">View YOLOv8 Detection model card on HuggingFace &rarr;</a>
+                </div>
             </div>
 
             <div id="yolov8-seg">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov8" target="_self">YOLOv8</a> &rsaquo; <strong>Segmentation</strong></p>
-                <h3>YOLOv8 Segmentation</h3>
-                <p><a href="https://huggingface.co/EdgeFirst/yolov8-seg">View on HuggingFace &rarr;</a></p>
+                <h3>YOLOv8 Segmentation &mdash; ONNX FP32 reference</h3>
+                <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX Det mAP</th><th>INT8 Det mAP</th><th>ONNX Mask mAP</th><th>INT8 Mask mAP</th></tr>
-                    <tr><td>Nano</td><td class="num">26.7%</td><td class="num">26.0%</td><td class="num">34.1%</td><td class="num">33.5%</td></tr>
+                    <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
+                    <tr><td>Nano</td><td class="num">50.04%</td><td class="num">35.28%</td><td class="num">46.03%</td><td class="num">27.41%</td></tr>
+                    <tr><td>Small</td><td class="num">58.85%</td><td class="num">43.05%</td><td class="num">53.98%</td><td class="num">32.55%</td></tr>
+                    <tr><td>Medium</td><td class="num">63.86%</td><td class="num">47.84%</td><td class="num">58.74%</td><td class="num">35.59%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov8-seg">View YOLOv8 Segmentation model card on HuggingFace &rarr;</a>
+                </div>
             </div>
             <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
         </div>
@@ -497,83 +580,28 @@
             <div class="model-grid">
                 <div class="model-card">
                     <h3><a href="#yolov5-det" target="_self">Detection</a></h3>
-                    <p class="meta">n/s/m/l/x &middot; Nano mAP@0.5: 49.6% &middot; <a href="https://huggingface.co/EdgeFirst/yolov5-det">HF Repo</a></p>
+                    <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 47.77% &middot; <a href="https://huggingface.co/EdgeFirst/yolov5-det">HF Repo</a></p>
                 </div>
             </div>
 
-            <h4>Size Scaling &mdash; Detection (ONNX FP32)</h4>
-            <table class="data-table">
-                <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th></tr>
-                <tr><td>Nano</td><td class="num">1.9M</td><td class="num">4.5</td><td class="num">49.6%</td><td class="num">33.0%</td></tr>
-                <tr><td>Small</td><td class="num">7.2M</td><td class="num">16.5</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Medium</td><td class="num">21.2M</td><td class="num">49.0</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>Large</td><td class="num">46.5M</td><td class="num">109.1</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-                <tr><td>XLarge</td><td class="num">86.7M</td><td class="num">205.7</td><td class="muted">&mdash;</td><td class="muted">&mdash;</td></tr>
-            </table>
-
             <div id="yolov5-det">
                 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov5" target="_self">YOLOv5</a> &rsaquo; <strong>Detection</strong></p>
-                <h3>YOLOv5 Detection</h3>
-                <p><a href="https://huggingface.co/EdgeFirst/yolov5-det">View on HuggingFace &rarr;</a></p>
+                <h3>YOLOv5 Detection &mdash; ONNX FP32 reference</h3>
+                <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
                 <table class="data-table">
-                    <tr><th>Size</th><th>ONNX FP32 mAP@0.5</th><th>TFLite INT8 mAP@0.5</th><th>INT8 Drop</th></tr>
-                    <tr><td>Nano</td><td class="num">49.6%</td><td class="num">46.2%</td><td class="num">−3.4 pp</td></tr>
+                    <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
+                    <tr><td>Nano</td><td class="num">1.9M</td><td class="num">4.5</td><td class="num">47.77%</td><td class="num">32.93%</td><td class="num">35.68%</td></tr>
+                    <tr><td>Small</td><td class="num">7.2M</td><td class="num">16.5</td><td class="num">57.32%</td><td class="num">41.27%</td><td class="num">44.90%</td></tr>
+                    <tr><td>Medium</td><td class="num">21.2M</td><td class="num">49.0</td><td class="num">63.29%</td><td class="num">47.00%</td><td class="num">51.39%</td></tr>
                 </table>
+                <div class="drill-down">
+                    <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
+                    <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov5-det">View YOLOv5 Detection model card on HuggingFace &rarr;</a>
+                </div>
             </div>
             <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
         </div>
 
-        <!-- ═══════════════════════════════════════════════════════ -->
-        <!-- WORKFLOW & VALIDATION                                  -->
-        <!-- ═══════════════════════════════════════════════════════ -->
-
-        <h2>Validation Pipeline</h2>
-        <div class="diagram-container">
-            <img src="01-ecosystem.png" alt="EdgeFirst Model Zoo Ecosystem">
-        </div>
-        <p>
-            Every artifact in the Model Zoo is measured on the <strong>same dataset on the same hardware</strong> users deploy on. <a href="https://edgefirst.studio"><strong>EdgeFirst Studio</strong></a> manages datasets, training, multi-format export, and reference validation; on-target runs happen on a board farm of i.MX 8M Plus, i.MX 95, Ara-240, Hailo, and Jetson devices. Accuracy numbers and per-stage timing are pushed back to Studio session metrics and consumed by this Model Zoo when generating each model card.
-        </p>
-
-        <h3>End-to-end Flow</h3>
-        <div class="diagram-container">
-            <img src="02-model-lifecycle.png" alt="Model Lifecycle: 5 stages from training to publication">
-        </div>
-        <p>
-            Each training session produces a single set of weights. The export pipeline emits ONNX FP32, INT8 TFLite, and platform-specific compiled formats (i.MX 95 Neutron, NXP Ara-240 DVM, Hailo HEF, Jetson TensorRT). Every output is paired with an on-target validation run that captures both <strong>accuracy</strong> (COCO/LVIS mAP against the validation set) and <strong>full-pipeline timing</strong>. The ONNX FP32 run from each training session serves as the reference baseline; quantization and runtime loss are measured relative to it, not relative to externally-published numbers.
-        </p>
-
-        <h3>EdgeFirst Profiler</h3>
-        <div class="diagram-container">
-            <img src="03-on-target-validation.png" alt="On-Target Validation Pipeline">
-        </div>
-        <p>
-            The <strong>EdgeFirst Profiler</strong> is the on-target agent that drives every validation. Given a model and a dataset, it runs full inference on the target device, captures per-image predictions in the EdgeFirst Arrow/Parquet format, and emits a <a href="https://perfetto.dev"><strong>Perfetto</strong></a> trace alongside the predictions. The Profiler is hardware-aware: it loads each runtime through its native delegate &mdash; Verisilicon VX Delegate on i.MX 8M Plus, eIQ Neutron Delegate on i.MX 95, Kinara SDK on Ara-240, HailoRT on RPi5 + Hailo, TensorRT on Jetson &mdash; so every stage timed by the trace corresponds to what a deployed application would experience.
-        </p>
-
-        <h3>EdgeFirst Validator</h3>
-        <p>
-            The <strong>EdgeFirst Validator</strong> is the off-target post-processor. It consumes the Profiler's predictions and Perfetto trace, computes the full 12-metric COCO accuracy tuple via <code>pycocotools</code> (or <code>lvis-api</code> for large-vocabulary datasets), and rebuilds per-stage timing summaries from the trace. Results land in a structured YAML payload attached to the Studio validation session &mdash; the same payload the Model Zoo reads to render this page. Accuracy and timing are computed independently of the runtime that produced the predictions, so toolchain regressions surface as cross-platform divergence rather than silent failures.
-        </p>
-
-        <h3>EdgeFirst HAL</h3>
-        <p>
-            The <a href="https://github.com/EdgeFirstAI/hal"><strong>EdgeFirst Hardware Abstraction Layer</strong> (HAL)</a> provides the hardware-accelerated primitives used at both validation and deployment time. The Profiler uses HAL for letterbox resize, color-space conversion, normalization, layout conversion, and post-decode (YOLO/ModelPack output decoding, NMS, mask materialisation). HAL automatically selects DMA-BUF, OpenGL ES, NXP G2D, or CPU paths depending on the platform &mdash; so the timing measured during validation reflects the same accelerated path a production runtime would take. HAL ships as a Rust crate, a Python package, and a C library under Apache 2.0.
-        </p>
-
-        <h3>Latency and Pipelined Throughput</h3>
-        <p>
-            Two complementary timing surfaces are reported per validation:
-        </p>
-        <table class="data-table">
-            <tr><th>Surface</th><th>What it captures</th><th>When it&#39;s present</th></tr>
-            <tr><td><code>timing.inline</code></td><td>Per-image <code>preprocess_ms</code>, <code>inference_ms</code>, <code>postprocess_ms</code> with min / mean / median / p95 / p99 / max</td><td>Always &mdash; the universal contract every producer fills in</td></tr>
-            <tr><td><code>timing.trace</code></td><td>Full per-stage breakdown from the Perfetto trace (typically 25&ndash;33 stages including delegate work, tensor moves, decode passes, NMS, parquet flush), plus end-to-end FPS distribution</td><td>When the Profiler emits a sidecar trace (almost all runs)</td></tr>
-        </table>
-        <p>
-            <strong>Throughput exceeds the sum of stage latencies</strong> because the runtime pipelines I/O, host preprocessing, NPU inference, and decode across frames. Reporting <code>1000 / (preprocess + inference + postprocess)</code> understates real throughput; the Model Zoo uses the <strong>measured end-to-end FPS</strong> from the trace (<code>trace.fps.median</code>) as the headline number. As a concrete example, YOLOv5n on i.MX 95 Neutron has per-stage means 21.7 ms preprocess + 12.2 ms inference + 15.8 ms postprocess (naive estimate ~20 FPS), but the measured pipelined throughput is <strong>56 FPS median</strong> &mdash; the 2.8&times; gap is the value the pipelining delivers.
-        </p>
-
         <!-- ═══════════════════════════════════════════════════════ -->
         <!-- NAMING & VARIANTS                                      -->
         <!-- ═══════════════════════════════════════════════════════ -->