README.md

---
language:
- en
license: apache-2.0
library_name: pytorch
tags:
- safetensors
- janegpt
- janus
- react
- interactive
- vite
- intent-classification
- transformer
- virtual-assistant
- nlp
- voice-assistant
- offline-ai
- edge-deployment
- nlu
- transformers
- slot-filling
- multitask-learning
- assistant-runtime
demo: https://janus-web-demo.vercel.app/
pipeline_tag: text-classification
model-index:
- name: JaneGPT-v2-Janus
  results:
  - task:
      type: text-classification
    dataset:
      name: BANKING77
      type: mteb/banking77
    metrics:
    - type: precision
      value: 1.0
      name: OOD Precision
    - type: f1
      value: 0.878
      name: OOD F1
    - type: recall
      value: 0.7825
      name: OOD Recall
  - task:
      type: text-classification
    dataset:
      name: CLINC OOS
      type: clinc/clinc_oos
    metrics:
    - type: precision
      value: 1.0
      name: OOD Precision
    - type: f1
      value: 0.7923
      name: OOD F1
    - type: recall
      value: 0.656
      name: OOD Recall
  - task:
      type: text-classification
    metrics:
    - type: accuracy
      value: 0.9983
      name: Validation Domain Accuracy
    - type: accuracy
      value: 0.9987
      name: Validation Action Accuracy
    - type: f1
      value: 1.0
      name: Slot Extraction F1
---

# JaneGPT v2 Janus - Intent Classification Model

---

![Total Downloads](https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fhuggingface.co%2Fapi%2Fmodels%2FRavinduSen%2FJaneGPT-v2-Janus%3Fexpand%3DdownloadsAllTime&query=downloadsAllTime&label=Total%20Downloads&color=blue)
![Monthly Downloads](https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fhuggingface.co%2Fapi%2Fmodels%2FRavinduSen%2FJaneGPT-v2-Janus&query=downloads&label=Downloads%2FMonth&color=green)

---

<p align="center">
  <a href="https://janus-web-demo.vercel.app/">
    <img src="https://img.shields.io/badge/Launch-Temple_of_Janus-6366f1?style=for-the-badge&logo=vercel" alt="Experience Janus">
  </a>
</p>

<style>
  /* ---- Hero section ---- */
  .hero-banner { text-align: center; margin-bottom: 12px; }
  .hero-banner img { max-width: 100%; border-radius: 12px; }

  .subtitle {
    text-align: center;
    font-size: 1.15em;
    color: #6b7a90;
    margin-bottom: 28px;
  }

  /* ---- Stat cards (mini key-value row) ---- */
  .stat-row {
    display: flex;
    flex-wrap: wrap;
    gap: 14px;
    justify-content: center;
    margin-bottom: 32px;
  }
  .stat-card {
    background: #0f1923;
    border: 1px solid #1e3048;
    border-radius: 10px;
    padding: 14px 22px;
    min-width: 150px;
    text-align: center;
    cursor: default;
  }
  .stat-card .stat-value {
    font-size: 1.65em;
    font-weight: 700;
    color: #58a6ff;
    line-height: 1.15;
  }
  .stat-card .stat-label {
    font-size: 0.82em;
    color: #8b9ab5;
    margin-top: 4px;
    text-transform: uppercase;
    letter-spacing: 0.5px;
  }

  /* ---- Section headings ---- */
  .section-heading {
    font-size: 1.35em;
    font-weight: 700;
    color: #e6edf3;
    border-bottom: 1px solid #1e3048;
    padding-bottom: 6px;
    margin-top: 36px;
    margin-bottom: 16px;
  }

  /* ---- Horizontal bar chart (interactive) ---- */
  .chart-container {
    background: #0d1520;
    border: 1px solid #1a2a3e;
    border-radius: 12px;
    padding: 24px 28px 60px 28px;
    margin-bottom: 24px;
    display: flex;
    flex-direction: column;
  }
  .chart-title {
    font-size: 1.1em;
    font-weight: 700;
    color: #e2e8f0;
    margin-bottom: 4px;
  }
  .chart-subtitle {
    font-size: 0.88em;
    color: #7a8ba3;
    margin-bottom: 14px;
    line-height: 1.35;
    position: relative;
    z-index: 3;
  }
  .bar-group {
    margin-bottom: 18px;
  }
  .bar-label {
    font-size: 0.92em;
    color: #c5d0de;
    margin-bottom: 5px;
    font-weight: 600;
  }

  /* NEW — replace with this */
  .bar-group {
    display: flex;
    flex-direction: column;
    align-items: center;
    flex: 1;
  }
  .bar-label {
    font-size: 0.78em;
    color: #8ea2b8;
    margin-top: 6px;
    text-align: center;
    font-weight: 600;
  }
  .bar-track {
    background: #162030;
    border-radius: 8px 8px 0 0;
    width: 100%;
    height: 120px;
    position: relative;
    display: flex;
    align-items: flex-end;
    overflow: hidden;
  }
  .bar-fill {
    width: 100%;
    border-radius: 6px 6px 0 0;
    transition: height 0.8s cubic-bezier(.22,.61,.36,1), box-shadow 0.3s ease;
    position: relative;
    min-height: 24px;
    display: flex;
    align-items: flex-start;
    justify-content: center;
  }
  .bar-fill:hover {
    box-shadow: 0 0 20px rgba(59, 130, 246, 0.3), inset 0 0 0 999px rgba(255,255,255,0.06);
  }
  .bar-value {
    position: absolute;
    top: 7px;
    left: 50%;
    transform: translateX(-50%);
    font-size: 0.78em;
    font-weight: 700;
    color: #000000;
    white-space: nowrap;
    line-height: 1;
    text-shadow: 0 1px 2px rgba(0,0,0,0.45);
    z-index: 2;
    pointer-events: none;
  }
  .bar-note {
    font-size: 0.78em;
    color: #5c6d82;
    margin-top: 6px;
    font-style: italic;
  }

  /* NEW — add this */
  .charts-row {
    display: flex;
    gap: 16px;
    align-items: stretch;
    margin-bottom: 24px;
  }
  .chart-container-sm {
    background: #0d1520;
    border: 1px solid #1a2a3e;
    border-radius: 12px;
    padding: 18px 16px;
    flex: 1;
    display: flex;
    flex-direction: column;
  }
  .bars-row {
    display: flex;
    gap: 8px;
    align-items: flex-end;
    min-height: 150px;
    height: 150px;
    border-bottom: 1px solid #1e3048;
    margin-bottom: 8px;
    margin-top: 10px;
  }
  .bars-row-wide {
    gap: 16px;
  }
  .ood-filter-row {
    margin-bottom: 14px;
    position: relative;
    z-index: 4;
  }
  .chart-bars {
    margin-top: 8px;
    position: relative;
    z-index: 1;
  }

  @media (max-width: 900px) {
    .charts-row {
      flex-direction: column;
      align-items: stretch;
    }
  }

  /* ---- Filter buttons (CSS-only via radio hack) ---- */
  .filter-row {
    display: flex;
    flex-wrap: wrap;
    gap: 8px;
    margin-bottom: 18px;
  }
  .filter-row input[type="radio"] {
    display: none;
  }
  input[name="ood-filter"],
  input[name="cm-toggle"] {
    position: absolute;
    opacity: 0;
    width: 0;
    height: 0;
    pointer-events: none;
  }
  .filter-btn {
    display: inline-block;
    padding: 5px 14px;
    font-size: 0.82em;
    font-weight: 600;
    border-radius: 20px;
    border: 1px solid #253649;
    background: #111c2a;
    color: #8ea2b8;
    cursor: pointer;
    transition: background 0.2s, color 0.2s, border-color 0.2s;
    user-select: none;
  }
  .filter-btn:hover {
    background: #1a2d42;
    color: #c5d8e8;
  }
  #ood-both:checked ~ .ood-filter-row label[for="ood-both"],
  #ood-banking77:checked ~ .ood-filter-row label[for="ood-banking77"],
  #ood-clinc:checked ~ .ood-filter-row label[for="ood-clinc"] {
    background: #1f6feb;
    color: #fff;
    border-color: #1f6feb;
  }
  /* Bar groups visibility controlled by radio state */
  .bar-group[data-dataset] { display: flex; }
  #ood-banking77:checked ~ .chart-bars .bar-group[data-dataset="clinc_oos"] { display: none; }
  #ood-clinc:checked ~ .chart-bars .bar-group[data-dataset="banking77"] { display: none; }
  #ood-both:checked ~ .chart-bars .bar-group[data-dataset] { display: flex; }

  /* ---- Comparison table ---- */
  .comparison-table {
    width: 100%;
    border-collapse: collapse;
    font-size: 0.9em;
    margin-bottom: 16px;
  }
  .comparison-table th {
    background: #111c2a;
    color: #8ea2b8;
    font-weight: 700;
    text-transform: uppercase;
    font-size: 0.78em;
    letter-spacing: 0.6px;
    padding: 10px 14px;
    border-bottom: 2px solid #1e3048;
    text-align: left;
  }
  .comparison-table td {
    padding: 9px 14px;
    border-bottom: 1px solid #152232;
    color: #c5d0de;
  }
  .comparison-table tr:hover td {
    background: #111e2e;
  }
  .tag-janus {
    display: inline-block;
    background: #1a3a2a;
    color: #4ade80;
    border-radius: 6px;
    padding: 1px 8px;
    font-size: 0.85em;
    font-weight: 600;
  }
  .tag-v2 {
    display: inline-block;
    background: #1a2a3a;
    color: #60a5fa;
    border-radius: 6px;
    padding: 1px 8px;
    font-size: 0.85em;
    font-weight: 600;
  }

  /* ---- Segmented stacked bar (single bar per panel) ---- */
  .stacked-bar-wrapper {
    margin-bottom: 12px;
    position: relative;
  }
  .stacked-bar-label {
    font-size: 0.85em;
    font-weight: 600;
    color: #8b9ab5;
    margin-bottom: 6px;
  }
  .stacked-bar {
    display: flex;
    height: 38px;
    border-radius: 8px;
    overflow: hidden;
    background: #162030;
    cursor: default;
  }
  .stacked-seg {
    height: 100%;
    transition: filter 0.2s ease, flex 0.3s ease;
    position: relative;
  }
  .stacked-seg:hover {
    filter: brightness(1.4) saturate(1.3);
    outline: 2px solid #fff;
    outline-offset: -2px;
    z-index: 5;
  }
  /* Tooltip for stacked bar segments */
  .stacked-tooltip {
    display: none;
    position: absolute;
    left: 50%;
    transform: translateX(-50%);
    top: calc(100% + 6px);
    background: #0f1923;
    border: 1px solid #253649;
    border-radius: 8px;
    padding: 8px 14px;
    font-size: 0.82em;
    color: #e2e8f0;
    z-index: 20;
    white-space: nowrap;
    box-shadow: 0 6px 20px rgba(0,0,0,0.5);
    pointer-events: none;
  }
  .stacked-tooltip::before {
    content: '';
    position: absolute;
    top: -6px;
    left: 50%;
    transform: translateX(-50%);
    border-left: 6px solid transparent;
    border-right: 6px solid transparent;
    border-bottom: 6px solid #253649;
  }
  .stacked-seg:hover .stacked-tooltip {
    display: block;
  }
  .stacked-tooltip .st-name { font-weight: 700; color: #4ade80; }
  .stacked-tooltip .st-val { color: #8ea2b8; }

  /* ---- Confusion matrix panel toggle ---- */
  #cm-domain:checked ~ .confusion-toggle label[for="cm-domain"],
  #cm-action:checked ~ .confusion-toggle label[for="cm-action"] {
    background: #1f6feb;
    color: #fff;
    border-color: #1f6feb;
  }
  .confusion-toggle {
    display: flex;
    gap: 8px;
    margin-bottom: 18px;
  }
  .cm-panel[data-cm="domain"] { display: block; }
  .cm-panel[data-cm="action"] { display: none; }
  #cm-action:checked ~ .cm-panel[data-cm="domain"] { display: none; }
  #cm-action:checked ~ .cm-panel[data-cm="action"] { display: block; }
  #cm-domain:checked ~ .cm-panel[data-cm="domain"] { display: block; }
  #cm-domain:checked ~ .cm-panel[data-cm="action"] { display: none; }

  /* ---- Confusion stacked bar ---- */
  .cm-stacked-bar {
    display: flex;
    height: 44px;
    border-radius: 10px;
    overflow: visible;
    background: #162030;
    position: relative;
    margin-bottom: 8px;
  }
  .cm-seg {
    height: 100%;
    position: relative;
    transition: filter 0.2s ease;
    cursor: default;
    /* thin separator between segments */
    border-right: 1px solid rgba(0,0,0,0.35);
  }
  .cm-seg:last-child { border-right: none; }
  .cm-seg:hover {
    filter: brightness(1.45) saturate(1.3);
    z-index: 10;
  }
  /* Tooltip on hover */
  .cm-seg .cm-tip {
    display: none;
    position: absolute;
    bottom: calc(100% + 8px);
    left: 50%;
    transform: translateX(-50%);
    background: #0f1923;
    border: 1px solid #253649;
    border-radius: 8px;
    padding: 10px 16px;
    font-size: 0.82em;
    color: #e2e8f0;
    z-index: 30;
    white-space: nowrap;
    box-shadow: 0 6px 24px rgba(0,0,0,0.55);
    pointer-events: none;
  }
  .cm-seg .cm-tip::after {
    content: '';
    position: absolute;
    bottom: -6px;
    left: 50%;
    transform: translateX(-50%);
    border-left: 6px solid transparent;
    border-right: 6px solid transparent;
    border-top: 6px solid #253649;
  }
  .cm-seg:hover .cm-tip { display: block; }
  .cm-tip .tip-name { font-weight: 700; color: #58a6ff; }
  .cm-tip .tip-samples { color: #4ade80; }
  .cm-tip .tip-acc { color: #fbbf24; }
  .cm-tip .tip-miss { color: #f87171; }
  .cm-legend {
    display: flex;
    flex-wrap: wrap;
    gap: 10px 18px;
    margin-top: 10px;
    margin-bottom: 6px;
  }
  .cm-legend-item {
    display: flex;
    align-items: center;
    gap: 6px;
    font-size: 0.78em;
    color: #8b9ab5;
  }
  .cm-legend-swatch {
    width: 12px;
    height: 12px;
    border-radius: 3px;
    display: inline-block;
  }

  /* ---- Code blocks ---- */
  details {
    background: #0d1520;
    border: 1px solid #1a2a3e;
    border-radius: 8px;
    margin-bottom: 10px;
    padding: 0;
  }
  details summary {
    cursor: pointer;
    padding: 12px 18px;
    font-weight: 600;
    color: #c5d0de;
    list-style: none;
  }
  details summary::-webkit-details-marker { display: none; }
  details summary::before { content: "▸ "; color: #58a6ff; }
  details[open] summary::before { content: "▾ "; }
  details > div {
    padding: 0 18px 16px;
  }
</style>

<!-- ===== Hero ===== -->
<div class="hero-banner">
  <img src="assets/jane-janus-glitch.webp" alt="Jane Janus animated hero banner" width="980" />
</div>

<p class="subtitle">Hierarchical command understanding with state-aware runtime behavior for practical assistant workflows.</p>

<!-- ===== Stat Cards ===== -->
<div class="stat-row">
  <div class="stat-card">
    <div class="stat-value">7.95M</div>
    <div class="stat-label">Parameters</div>
  </div>
  <div class="stat-card">
    <div class="stat-value">82</div>
    <div class="stat-label">Runtime Turns</div>
  </div>
  <div class="stat-card">
    <div class="stat-value">0</div>
    <div class="stat-label">Errors</div>
  </div>
  <div class="stat-card">
    <div class="stat-value">25.3 ms</div>
    <div class="stat-label">Mean Latency</div>
  </div>
  <div class="stat-card">
    <div class="stat-value">100%</div>
    <div class="stat-label">OOD Precision</div>
  </div>
  <div class="stat-card">
    <div class="stat-value">30.6 MB</div>
    <div class="stat-label">Checkpoint</div>
  </div>
</div>

---

## 🏛️ The Temple of Janus (Web Experience)

We have deployed a dedicated interactive environment to showcase the essence of JaneGPT-v2 Janus. 

**Note:** This is a visual and technical walkthrough; it does not feature a live chat interface.

* **[🔗 Enter the Experience](https://janus-web-demo.vercel.app/)**
* **Best Viewed On:** Desktop (Chrome/Edge) for full hardware-accelerated 3D effects.

---

## Quickstart (2 minutes)

<details open>
<summary><strong>Install + first prediction</strong></summary>
<div>

```bash
pip install -r requirements.txt
```

```python
from janegpt_v2_janus.inference import JaneGPTv3NLU

nlu = JaneGPTv3NLU(
    model_path="weights/janegpt_v2_janus.pt",
    tokenizer_path="weights/tokenizer.json",
)

state = {}
result = nlu.predict("set volume", state=state)
print(result)

if result.get("type") == "command":
    state = nlu.update_state(result, state)
```

</div>
</details>

<details>
<summary><strong>Runtime wrapper (recommended for assistant flows)</strong></summary>
<div>

```python
from runtime.jane_nlu_runtime import JaneNLURuntime

rt = JaneNLURuntime(base_dir=".")
state = {}

out, state = rt.handle_turn("set volume", state)
print(out)  # expected: clarify prompt for missing VALUE

out, state = rt.handle_turn("55", state)
print(out)  # expected: resolved local command
```

</div>
</details>

<details>
<summary><strong>Run bundled demos</strong></summary>
<div>

```bash
python examples/demo_inference.py
python examples/demo_runtime.py
python examples/demo_runtime_suite.py
```

</div>
</details>

---

## What You Get

- Single-pass multitask prediction: domain + action + BIO slots.
- Runtime-safe clarification loops for missing required slots.
- Stateful follow-ups (for example, "that is not enough" after a volume change).
- Local command routing with controlled chat fallback.
- Compact deployment footprint: ~30.62 MB checkpoint.

---

## Model Architecture

### Interactive Architecture Visualization

<style>
  /* ---- Architecture Visualization ---- */
  .arch-container {
    background: linear-gradient(135deg, #0a0f1a 0%, #0d1520 100%);
    border: 1px solid #1a2a3e;
    border-radius: 12px;
    padding: 28px;
    margin-bottom: 24px;
    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
  }

  .arch-flow {
    display: flex;
    flex-direction: column;
    gap: 12px;
  }

  .arch-layer {
    background: #0d1520;
    border: 2px solid #1e3048;
    border-radius: 10px;
    padding: 18px;
    cursor: pointer;
    transition: all 0.3s ease;
    position: relative;
    overflow: hidden;
  }

  .arch-layer::before {
    content: '';
    position: absolute;
    top: 0;
    left: 0;
    right: 0;
    height: 2px;
    background: linear-gradient(90deg, #3b82f6, #8b5cf6, #ec4899, transparent);
    opacity: 0;
    transition: opacity 0.3s ease;
  }

  .arch-layer:hover {
    background: #111c2e;
    border-color: #2d5aa6;
    box-shadow: 0 8px 24px rgba(59, 130, 246, 0.15);
  }

  .arch-layer:hover::before {
    opacity: 1;
  }

  .arch-layer-title {
    font-size: 1.05em;
    font-weight: 700;
    color: #58a6ff;
    margin-bottom: 10px;
    display: flex;
    align-items: center;
    justify-content: center;
    gap: 8px;
    text-align: center;
  }

  .arch-layer-icon {
    display: none;
  }

  .arch-layer-desc {
    font-size: 0.9em;
    color: #8b9ab5;
    line-height: 1.5;
    max-height: 0;
    overflow: hidden;
    transition: max-height 0.3s ease;
  }

  .arch-layer.expanded .arch-layer-desc {
    max-height: 500px;
  }

  .arch-layer-content {
    display: grid;
    grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
    gap: 12px;
    margin-top: 12px;
    max-height: 0;
    overflow: hidden;
    transition: max-height 0.3s ease;
  }

  .arch-layer.expanded .arch-layer-content {
    max-height: 600px;
  }

  .arch-spec-item {
    background: #162030;
    border: 1px solid #253649;
    border-radius: 8px;
    padding: 10px 12px;
    font-size: 0.82em;
    text-align: center;
  }

  .arch-spec-label {
    color: #4ade80;
    font-weight: 600;
    display: block;
    margin-bottom: 4px;
  }

  .arch-spec-value {
    color: #c5d0de;
    font-family: 'Monaco', 'Courier New', monospace;
  }

  .arch-arrow {
    text-align: center;
    color: #4ade80;
    font-size: 1.2em;
    padding: 4px 0;
    opacity: 0.6;
  }

  .arch-tasks {
    display: grid;
    grid-template-columns: repeat(auto-fit, minmax(180px, 1fr));
    gap: 12px;
    margin-top: 12px;
  }

  .arch-task {
    background: linear-gradient(135deg, #1a2f4a 0%, #0f1f38 100%);
    border: 1px solid #253649;
    border-left: 3px solid #3b82f6;
    border-radius: 8px;
    padding: 12px;
    font-size: 0.85em;
  }

  .arch-task.domain {
    border-left-color: #3b82f6;
  }

  .arch-task.action {
    border-left-color: #8b5cf6;
  }

  .arch-task.slot {
    border-left-color: #ec4899;
  }

  .arch-task-name {
    font-weight: 700;
    color: #e2e8f0;
    margin-bottom: 6px;
    text-align: center;
  }

  .arch-task-detail {
    color: #8b9ab5;
    font-size: 0.78em;
    line-height: 1.4;
    text-align: center;
  }

  .arch-loss-box {
    background: #1a2a3a;
    border: 1px solid #1e3a4f;
    border-radius: 8px;
    padding: 14px;
    margin-top: 12px;
    font-family: 'Monaco', 'Courier New', monospace;
    font-size: 0.82em;
    color: #4ade80;
    overflow-x: auto;
  }

  .arch-decision {
    background: #1a3a2a;
    border-left: 3px solid #10b981;
    border-radius: 6px;
    padding: 10px;
    margin-top: 8px;
    font-size: 0.82em;
    color: #a7f3d0;
  }

  @media (max-width: 900px) {
    .arch-layer-content {
      grid-template-columns: 1fr;
    }
    
    .arch-tasks {
      grid-template-columns: 1fr;
    }
  }
</style>

<div class="arch-container">

  <div class="arch-layer" onclick="this.classList.toggle('expanded')">
    <div class="arch-layer-title">
      <span class="arch-layer-icon">🔤</span>
      1. Tokenization & Embedding Layer
    </div>
    <div class="arch-layer-desc">
      Input text is converted to token IDs and projected into a 256-dimensional embedding space.
    </div>
    <div class="arch-layer-content">
      <div class="arch-spec-item">
        <span class="arch-spec-label">Tokenizer</span>
        <span class="arch-spec-value">BPE, vocab=8,192</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Max Length</span>
        <span class="arch-spec-value">96 tokens</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Output Shape</span>
        <span class="arch-spec-value">(batch, 96, 256)</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Embedding</span>
        <span class="arch-spec-value">8192 → 256 dim</span>
      </div>
    </div>
  </div>

  <div class="arch-arrow">↓</div>

  <div class="arch-layer" onclick="this.classList.toggle('expanded')">
    <div class="arch-layer-title">
      <span class="arch-layer-icon"></span>
      2. Transformer Backbone (8 Blocks)
    </div>
    <div class="arch-layer-desc">
      Bidirectional attention layers with residual connections. Each block processes hidden states through grouped query attention and feed-forward networks.
    </div>
    <div class="arch-layer-content">
      <div class="arch-spec-item">
        <span class="arch-spec-label">Attention Type</span>
        <span class="arch-spec-value">Grouped Query (GQA)</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Query Heads</span>
        <span class="arch-spec-value">8 heads</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">KV Heads</span>
        <span class="arch-spec-value">4 heads (2:1)</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Head Dimension</span>
        <span class="arch-spec-value">32 (256÷8)</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Position Embedding</span>
        <span class="arch-spec-value">RoPE</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">FFN Expansion</span>
        <span class="arch-spec-value">256 → 672 → 256</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">FFN Activation</span>
        <span class="arch-spec-value">SwiGLU</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Normalization</span>
        <span class="arch-spec-value">RMSNorm</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Causal Masking</span>
        <span class="arch-spec-value">OFF (bidirectional)</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Dropout Rate</span>
        <span class="arch-spec-value">0.1</span>
      </div>
    </div>
    <div class="arch-decision">
      Grouped Query Attention reduces KV cache 50% while maintaining quality
    </div>
  </div>

  <div class="arch-arrow">↓</div>

  <div class="arch-layer" onclick="this.classList.toggle('expanded')">
    <div class="arch-layer-title">
      <span class="arch-layer-icon"></span>
      3. Multi-Task Prediction Heads (Parallel)
    </div>
    <div class="arch-layer-desc">
      Three independent classification heads process the backbone output simultaneously for domain, action, and slot predictions.
    </div>
    <div class="arch-tasks">
      <div class="arch-task domain">
        <div class="arch-task-name">Domain Head</div>
        <div class="arch-task-detail">
          <strong>Input:</strong> Last token (pooled)<br/>
          <strong>Arch:</strong> Linear(256) → GELU → Dropout → Linear(10)<br/>
          <strong>Output:</strong> 10 classes
        </div>
      </div>
      <div class="arch-task action">
        <div class="arch-task-name">Action Head</div>
        <div class="arch-task-detail">
          <strong>Input:</strong> Last token (pooled)<br/>
          <strong>Arch:</strong> Linear(256) → GELU → Dropout → Linear(33)<br/>
          <strong>Output:</strong> 33 classes
        </div>
      </div>
      <div class="arch-task slot">
        <div class="arch-task-name">Slot Head</div>
        <div class="arch-task-detail">
          <strong>Input:</strong> All tokens<br/>
          <strong>Arch:</strong> Linear(256) → Linear(15 BIO)<br/>
          <strong>Output:</strong> 15 labels/token
        </div>
      </div>
    </div>
  </div>

  <div class="arch-arrow">↓</div>

  <div class="arch-layer" onclick="this.classList.toggle('expanded')">
    <div class="arch-layer-title">
      <span class="arch-layer-icon"></span>
      4. Output & Post-Processing
    </div>
    <div class="arch-layer-desc">
      Raw logits are converted to predictions. For slots, BIO tags are decoded into semantic spans.
    </div>
    <div class="arch-layer-content">
      <div class="arch-spec-item">
        <span class="arch-spec-label">Domain Output</span>
        <span class="arch-spec-value">10 classes</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Action Output</span>
        <span class="arch-spec-value">33 classes</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Slots Decoder</span>
        <span class="arch-spec-value">BIO → Spans</span>
      </div>
      <div class="arch-spec-item">
        <span class="arch-spec-label">Confidence</span>
        <span class="arch-spec-value">Softmax scores</span>
      </div>
    </div>
  </div>

</div>

### Training Objective

<div class="arch-container">
  <div class="arch-layer expanded">
    <div class="arch-layer-title">
      <span class="arch-layer-icon"></span>
      Weighted Multi-Task Loss
    </div>
    <div class="arch-loss-box">
loss = 1.0 × L_domain + 1.0 × L_action + 1.5 × L_slots

Where:
  L_domain = CrossEntropy(domain_logits, domain_labels)
  L_action = CrossEntropy(action_logits, action_labels)
  L_slots  = CrossEntropy(slot_logits, slot_labels)
             with ignore_index=-100 (padding)
    </div>
    <div class="arch-decision">
      Slot weight (1.5×) reflects higher complexity of sequence tagging vs. classification
    </div>
  </div>
</div>

### Architecture Specifications

| Component | Configuration | Details |
|-----------|---|---|
| **Backbone Type** | Transformer (GPT-style) | Bidirectional, non-causal attention |
| **Vocabulary Size** | 8,192 | BPE tokenization |
| **Embedding Dim** | 256 | Token + Rotary Position embeddings |
| **Attention Heads** | 8 Query, 4 KV | Grouped Query Attention (GQA) for efficiency |
| **Head Dimension** | 32 | per head_dim = embed_dim / num_heads |
| **Transformer Blocks** | 8 Layers | Each with Attn + FFN + Residuals |
| **Feed-Forward Hidden** | 672 | SwiGLU gate activation |
| **Position Encoding** | RoPE | Rotary Position Embeddings (theta=10000) |
| **Normalization** | RMSNorm | Pre-layer normalization |
| **Max Sequence Length** | 96 tokens | Approximately 60-80 words |
| **Dropout Rate** | 0.1 | Applied during training |
| **Total Parameters** | 7,949,626 | All trainable |
| **Parameter Breakdown** | Backbone: 7.80M, Task Heads: 146K | Efficient multitask design |

### Task Configuration

| Task | Type | Classes | Architecture |
|------|---|---|---|
| **Domain Classification** | Sequence-level | 10 domains | Pooled → Linear(256) → GELU → Linear(10) |
| **Action Classification** | Sequence-level | 33 actions | Pooled → Linear(256) → GELU → Linear(33) |
| **Slot Tagging** | Token-level | 15 BIO labels | Per-token → Linear(256) → Linear(15) |

---

## Benchmark Results
<div class="charts-row">

  <div class="chart-container-sm">
    <div class="chart-title">Runtime reliability</div>
    <div class="chart-subtitle">82-turn suite</div>
    <div class="bars-row">
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:100%; background:linear-gradient(0deg,#3b82f6,#60a5fa);"><span class="bar-value">82</span></div></div>
        <div class="bar-label">Turns</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:81.7%; background:linear-gradient(0deg,#8b5cf6,#a78bfa);"><span class="bar-value">67</span></div></div>
        <div class="bar-label">Local</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:14.6%; background:linear-gradient(0deg,#ec4899,#f472b6);"><span class="bar-value">12</span></div></div>
        <div class="bar-label">Clarify</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:1%; background:linear-gradient(0deg,#9ca3af,#d1d5db);"><span class="bar-value">0</span></div></div>
        <div class="bar-label">Errors</div>
      </div>
    </div>
    <div class="bar-note">fair_benchmarks.json</div>
  </div>

  <div class="chart-container-sm">
    <div class="chart-title">Predict latency</div>
    <div class="chart-subtitle">CUDA · batch=1 · lower is better</div>
    <div class="bars-row">
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:63.3%; background:linear-gradient(0deg,#3b82f6,#60a5fa);"><span class="bar-value">25.3ms</span></div></div>
        <div class="bar-label">P·mean</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:86.5%; background:linear-gradient(0deg,#8b5cf6,#a78bfa);"><span class="bar-value">34.6ms</span></div></div>
        <div class="bar-label">P·p95</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:88.4%; background:linear-gradient(0deg,#ec4899,#f472b6);"><span class="bar-value">35.4ms</span></div></div>
        <div class="bar-label">Fwd·mean</div>
      </div>
      <div class="bar-group">
        <div class="bar-track"><div class="bar-fill" style="height:91.8%; background:linear-gradient(0deg,#a78bfa,#c4b5fd);"><span class="bar-value">36.7ms</span></div></div>
        <div class="bar-label">Fwd·p95</div>
      </div>
    </div>
    <div class="bar-note">janus_model_report.json</div>
  </div>

</div>

<div class="chart-container">
  <div class="chart-title">OOD rejection quality</div>
  <div class="chart-subtitle">Schema-agnostic · hover values</div>

  <input type="radio" name="ood-filter" id="ood-both" checked />
  <input type="radio" name="ood-filter" id="ood-banking77" />
  <input type="radio" name="ood-filter" id="ood-clinc" />
  <div class="ood-filter-row filter-row">
  <label class="filter-btn" for="ood-both">Both</label>
  <label class="filter-btn" for="ood-banking77">BANKING77</label>
  <label class="filter-btn" for="ood-clinc">CLINC</label>
  </div>
  <div class="chart-bars">
  <div class="bars-row bars-row-wide">
<div class="bar-group" data-dataset="banking77">
<div class="bar-track"><div class="bar-fill" style="height:87.8%; background:linear-gradient(0deg,#3b82f6,#60a5fa);"><span class="bar-value">87.8%</span></div></div>
<div class="bar-label">B77 F1</div>
</div>
<div class="bar-group" data-dataset="banking77">
<div class="bar-track"><div class="bar-fill" style="height:100%; background:linear-gradient(0deg,#ec4899,#f472b6);"><span class="bar-value">100%</span></div></div>
<div class="bar-label">B77 Prec</div>
</div>
<div class="bar-group" data-dataset="banking77">
<div class="bar-track"><div class="bar-fill" style="height:78.25%; background:linear-gradient(0deg,#8b5cf6,#a78bfa);"><span class="bar-value">78.3%</span></div></div>
<div class="bar-label">B77 Rec</div>
</div>
<div class="bar-group" data-dataset="clinc_oos">
<div class="bar-track"><div class="bar-fill" style="height:79.23%; background:linear-gradient(0deg,#3b82f6,#60a5fa);"><span class="bar-value">79.2%</span></div></div>
<div class="bar-label">CL F1</div>
</div>
<div class="bar-group" data-dataset="clinc_oos">
<div class="bar-track"><div class="bar-fill" style="height:100%; background:linear-gradient(0deg,#ec4899,#f472b6);"><span class="bar-value">100%</span></div></div>
<div class="bar-label">CL Prec</div>
</div>
<div class="bar-group" data-dataset="clinc_oos">
<div class="bar-track"><div class="bar-fill" style="height:65.6%; background:linear-gradient(0deg,#8b5cf6,#a78bfa);"><span class="bar-value">65.6%</span></div></div>
<div class="bar-label">CL Rec</div>
</div>
</div>
</div>
<div class="bar-note">fair_benchmarks.json</div>
</div>

</div>

### Comprehensive Benchmark Summary

<div class="chart-container">
  <div class="chart-title">Full Benchmark Evidence</div>
  <div class="chart-subtitle">All values from real holdout evaluations — no synthetic or inflated numbers</div>

  <table class="comparison-table">
    <thead>
      <tr>
        <th>Metric</th>
        <th>Detail</th>
        <th><span class="tag-v2">Jane v2</span></th>
        <th><span class="tag-janus">Janus</span></th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td><strong>Speed (mean latency)</strong></td>
        <td>CUDA, batch=1</td>
        <td>31.60 ms</td>
        <td><strong>25.31 ms</strong></td>
      </tr>
      <tr>
        <td><strong>Throughput</strong></td>
        <td>CUDA, single GPU</td>
        <td>32 pred/sec</td>
        <td>Stable across 82 turns, 0 errors</td>
      </tr>
      <tr>
        <td><strong>OOD F1</strong></td>
        <td>BANKING77</td>
        <td><strong>94.31%</strong></td>
        <td>87.80%</td>
      </tr>
      <tr>
        <td><strong>OOD F1</strong></td>
        <td>CLINC OOS</td>
        <td><strong>89.16%</strong></td>
        <td>79.23%</td>
      </tr>
      <tr>
        <td><strong>OOD Precision</strong></td>
        <td>BANKING77</td>
        <td>99.35%</td>
        <td><strong>100.00%</strong></td>
      </tr>
      <tr>
        <td><strong>OOD Precision</strong></td>
        <td>CLINC OOS</td>
        <td>99.14%</td>
        <td><strong>100.00%</strong></td>
      </tr>
      <tr>
        <td><strong>OOD Recall</strong></td>
        <td>BANKING77</td>
        <td><strong>89.75%</strong></td>
        <td>78.25%</td>
      </tr>
      <tr>
        <td><strong>OOD Recall</strong></td>
        <td>CLINC OOS</td>
        <td><strong>81.00%</strong></td>
        <td>65.60%</td>
      </tr>
      <tr>
        <td><strong>Validation Accuracy</strong></td>
        <td>Domain (best epoch)</td>
        <td>—</td>
        <td><strong>99.83%</strong></td>
      </tr>
      <tr>
        <td><strong>Validation Accuracy</strong></td>
        <td>Action (best epoch)</td>
        <td>—</td>
        <td><strong>99.87%</strong></td>
      </tr>
      <tr>
        <td><strong>Validation Accuracy</strong></td>
        <td>Domain+Action pair (best epoch)</td>
        <td>—</td>
        <td><strong>99.83%</strong></td>
      </tr>
      <tr>
        <td><strong>Slot Extraction F1</strong></td>
        <td>All 15 slot types</td>
        <td>—</td>
        <td><strong>1.000 (100%)</strong></td>
      </tr>
      <tr>
        <td><strong>Training Loss</strong></td>
        <td>Epoch 1 → 4</td>
        <td>—</td>
        <td>0.060 → 0.020 → 0.002 → 0.001</td>
      </tr>
      <tr>
        <td><strong>Validation Loss</strong></td>
        <td>Epoch 1 → 3</td>
        <td>—</td>
        <td>0.0153 → 0.0116 → 0.0115 (stable)</td>
      </tr>
      <tr>
        <td><strong>Runtime Reliability</strong></td>
        <td>82-turn conversation test</td>
        <td>—</td>
        <td><strong>0 errors, 0 crashes</strong></td>
      </tr>
      <tr>
        <td><strong>Domain Confusion</strong></td>
        <td>10 domains</td>
        <td>—</td>
        <td>99%+ per-domain, minimal cross-confusion</td>
      </tr>
      <tr>
        <td><strong>Action Confusion</strong></td>
        <td>33 actions</td>
        <td>—</td>
        <td>Perfect diagonal, no action commonly confused</td>
      </tr>
    </tbody>
  </table>
</div>

---

## Live Output Shapes (click to expand)

<details>
<summary><strong>Command output</strong></summary>
<div>

```json
{
  "type": "command",
  "domain": "apps",
  "action": "launch",
  "slots": {
    "APP_NAME": {
      "text": "chrome",
      "start": 5,
      "end": 11,
      "confidence": 0.999
    }
  },
  "confidence": 0.97,
  "route": "local"
}
```

</div>
</details>

<details>
<summary><strong>Clarification output</strong></summary>
<div>

```json
{
  "type": "clarify",
  "question": "What value should I set it to?",
  "debug": {
    "domain": "volume",
    "action": "set",
    "reason": "missing_VALUE"
  }
}
```

</div>
</details>

<details>
<summary><strong>Label schema</strong></summary>
<div>

- **Domains (10):** volume, brightness, media, apps, browser, productivity, screen, window, system, conversation
- **Actions (33):** up, down, set, mute, unmute, play, pause, next, previous, launch, close, switch, search, set_reminder, screenshot, read, explain, undo, quit, chat, minimize, maximize, restore, focus, copy, paste, cut, lock, sleep, wifi_on, wifi_off, bluetooth_on, bluetooth_off
- **Slot labels (BIO, 15):** VALUE, APP_NAME, QUERY, DURATION, TIME, WINDOW_NAME, TEXT

</div>
</details>

---

## Visual Benchmark Evidence

<p align="center">
  <img src="reports/loss_per_epoch.png" alt="Train and validation loss" width="860" />
</p>

<p align="center">
  <img src="reports/train_loss_smoothed.png" alt="Smoothed train loss" width="860" />
</p>

<p align="center">
  <img src="reports/val_slot_f1.png" alt="Validation slot F1" width="860" />
</p>

### Confusion Matrix — Interactive Breakdown

<!-- FIX: radio inputs must be SIBLINGS of both .confusion-toggle AND .cm-panel divs
     so the CSS sibling combinator (~) can reach them all from the same parent level -->
<div class="chart-container">
<div class="chart-title">Per-Class True vs Predicted</div>
<div class="chart-subtitle">Single stacked bar per head — segment width = sample ratio. Hover any segment for details.</div>

<input type="radio" name="cm-toggle" id="cm-domain" checked />
<input type="radio" name="cm-toggle" id="cm-action" />
<div class="confusion-toggle">
<label class="filter-btn" for="cm-domain">Domains (10)</label>
<label class="filter-btn" for="cm-action">Actions (33)</label>
</div>

<div class="cm-panel" data-cm="domain">

<div class="stacked-bar-label">Domain Sample Distribution &mdash; 3,110 total samples &mdash; hover each segment</div>
<div class="cm-stacked-bar">
  <div class="cm-seg" style="flex:430; background:#3b82f6; border-radius:10px 0 0 10px;"><div class="cm-tip"><span class="tip-name">volume</span><br/><span class="tip-samples">430 samples (13.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:250; background:#f59e0b;"><div class="cm-tip"><span class="tip-name">brightness</span><br/><span class="tip-samples">250 samples (8.0%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:340; background:#10b981;"><div class="cm-tip"><span class="tip-name">media</span><br/><span class="tip-samples">340 samples (10.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:250; background:#ef4444;"><div class="cm-tip"><span class="tip-name">apps</span><br/><span class="tip-samples">250 samples (8.0%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:120; background:#8b5cf6;"><div class="cm-tip"><span class="tip-name">browser</span><br/><span class="tip-samples">120 samples (3.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:340; background:#ec4899;"><div class="cm-tip"><span class="tip-name">productivity</span><br/><span class="tip-samples">340 samples (10.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:120; background:#14b8a6;"><div class="cm-tip"><span class="tip-name">screen</span><br/><span class="tip-samples">120 samples (3.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:340; background:#f97316;"><div class="cm-tip"><span class="tip-name">window</span><br/><span class="tip-samples">340 samples (10.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:580; background:#6366f1;"><div class="cm-tip"><span class="tip-name">system</span><br/><span class="tip-samples">580 samples (18.6%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:340; background:#06b6d4; border-radius:0 10px 10px 0;"><div class="cm-tip"><span class="tip-name">conversation</span><br/><span class="tip-samples">340 samples (10.9%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
</div>

<div class="cm-legend">
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#3b82f6;"></span>volume</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#f59e0b;"></span>brightness</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#10b981;"></span>media</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#ef4444;"></span>apps</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#8b5cf6;"></span>browser</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#ec4899;"></span>productivity</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#14b8a6;"></span>screen</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#f97316;"></span>window</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#6366f1;"></span>system</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#06b6d4;"></span>conversation</div>
</div>

<div class="bar-note">Source: validation set confusion matrix — segment widths proportional to sample count</div>
</div>

<div class="cm-panel" data-cm="action">

<div class="stacked-bar-label">Action Sample Distribution &mdash; 3,205 total samples &mdash; hover each segment</div>
<div class="cm-stacked-bar">
  <div class="cm-seg" style="flex:170; background:#3b82f6; border-radius:10px 0 0 10px;"><div class="cm-tip"><span class="tip-name">up</span><br/><span class="tip-samples">170 samples (5.3%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:165; background:#2563eb;"><div class="cm-tip"><span class="tip-name">down</span><br/><span class="tip-samples">165 samples (5.1%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:170; background:#1d4ed8;"><div class="cm-tip"><span class="tip-name">set</span><br/><span class="tip-samples">170 samples (5.3%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#7c3aed;"><div class="cm-tip"><span class="tip-name">mute</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#6d28d9;"><div class="cm-tip"><span class="tip-name">unmute</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#10b981;"><div class="cm-tip"><span class="tip-name">play</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#059669;"><div class="cm-tip"><span class="tip-name">pause</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#047857;"><div class="cm-tip"><span class="tip-name">next</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#065f46;"><div class="cm-tip"><span class="tip-name">previous</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#ef4444;"><div class="cm-tip"><span class="tip-name">launch</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#dc2626;"><div class="cm-tip"><span class="tip-name">close</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#b91c1c;"><div class="cm-tip"><span class="tip-name">switch</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#f59e0b;"><div class="cm-tip"><span class="tip-name">search</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#d97706;"><div class="cm-tip"><span class="tip-name">set_reminder</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#ec4899;"><div class="cm-tip"><span class="tip-name">screenshot</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#db2777;"><div class="cm-tip"><span class="tip-name">read</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#be185d;"><div class="cm-tip"><span class="tip-name">explain</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#14b8a6;"><div class="cm-tip"><span class="tip-name">undo</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#0d9488;"><div class="cm-tip"><span class="tip-name">quit</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#0f766e;"><div class="cm-tip"><span class="tip-name">chat</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#f97316;"><div class="cm-tip"><span class="tip-name">minimize</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#ea580c;"><div class="cm-tip"><span class="tip-name">maximize</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#c2410c;"><div class="cm-tip"><span class="tip-name">restore</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#6366f1;"><div class="cm-tip"><span class="tip-name">focus</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#4f46e5;"><div class="cm-tip"><span class="tip-name">copy</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#4338ca;"><div class="cm-tip"><span class="tip-name">paste</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#06b6d4;"><div class="cm-tip"><span class="tip-name">cut</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#0891b2;"><div class="cm-tip"><span class="tip-name">lock</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#0e7490;"><div class="cm-tip"><span class="tip-name">sleep</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#84cc16;"><div class="cm-tip"><span class="tip-name">wifi_on</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#65a30d;"><div class="cm-tip"><span class="tip-name">wifi_off</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#a855f7;"><div class="cm-tip"><span class="tip-name">bluetooth_on</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
  <div class="cm-seg" style="flex:90; background:#9333ea; border-radius:0 10px 10px 0;"><div class="cm-tip"><span class="tip-name">bluetooth_off</span><br/><span class="tip-samples">90 samples (2.8%)</span><br/><span class="tip-acc">Accuracy: 100%</span><br/><span class="tip-miss">Misclassified: 0</span></div></div>
</div>

<div class="cm-legend">
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#3b82f6;"></span>up</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#2563eb;"></span>down</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#1d4ed8;"></span>set</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#7c3aed;"></span>mute</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#6d28d9;"></span>unmute</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#10b981;"></span>play</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#059669;"></span>pause</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#047857;"></span>next</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#065f46;"></span>previous</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#ef4444;"></span>launch</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#dc2626;"></span>close</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#b91c1c;"></span>switch</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#f59e0b;"></span>search</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#d97706;"></span>set_reminder</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#ec4899;"></span>screenshot</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#db2777;"></span>read</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#be185d;"></span>explain</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#14b8a6;"></span>undo</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#0d9488;"></span>quit</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#0f766e;"></span>chat</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#f97316;"></span>minimize</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#ea580c;"></span>maximize</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#c2410c;"></span>restore</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#6366f1;"></span>focus</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#4f46e5;"></span>copy</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#4338ca;"></span>paste</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#06b6d4;"></span>cut</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#0891b2;"></span>lock</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#0e7490;"></span>sleep</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#84cc16;"></span>wifi_on</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#65a30d;"></span>wifi_off</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#a855f7;"></span>bluetooth_on</div>
  <div class="cm-legend-item"><span class="cm-legend-swatch" style="background:#9333ea;"></span>bluetooth_off</div>
</div>

<div class="bar-note">Source: validation set confusion matrix — segment widths proportional to sample count</div>
</div>

</div>

<details>
<summary><strong>View original confusion matrix images</strong></summary>
<div>

<p align="center">
  <img src="reports/confusion_domain_val.png" alt="Domain confusion matrix" width="860" />
</p>

<p align="center">
  <img src="reports/confusion_action_val.png" alt="Action confusion matrix" width="860" />
</p>

</div>
</details>

<details>
<summary><strong>Additional diagnostics</strong></summary>
<div>

<p align="center">
  <img src="reports/lr_schedule.png" alt="Learning rate schedule" width="860" />
</p>

<p align="center">
  <img src="reports/epoch_time.png" alt="Epoch time profile" width="860" />
</p>

<p align="center">
  <img src="reports/train_loss_raw.png" alt="Raw training loss" width="860" />
</p>

</div>
</details>

---

## Upload-Ready Layout

```text
.
|- README.md
|- .gitattributes
|- LICENSE
|- requirements.txt
|- assets/
|  |- jane-janus-glitch.webp
|- janegpt_v2_janus/
|  |- __init__.py
|  |- architecture.py
|  |- dataset.py
|  |- inference.py
|  |- labels.py
|  |- multitask.py
|- runtime/
|  |- jane_nlu_runtime.py
|- examples/
|  |- demo_inference.py
|  |- demo_runtime.py
|  |- demo_runtime_suite.py
|- weights/
|  |- janegpt_v2_janus.pt
|  |- tokenizer.json
|- reports/
|  |- fair_benchmarks.json
|  |- fair_benchmarks.md
|  |- janus_model_report.json
|  |- janus_model_report.md
|  |- public_benchmarks.json
|  |- *.png benchmark visuals
```

---

## Limitations

- English-focused command language.
- Command NLU model, not an open-domain generative chatbot.
- MASSIVE and SNIPS mapped-intent accuracy is excluded from headline claims because mapping coverage is partial.

---

## Use Cases

- Virtual assistant command routing
- Smart home intent classification
- Voice command understanding
- Chatbot intent detection
- Edge device deployment (small enough for embedded systems)

---

## Part of the JANE Project

**JANE** — a fully 
offline, privacy-first AI voice assistant.

🔗 [JANE AI Assistant on GitHub](https://github.com/Ravindu-S/JANE-AI-Assistant)
🔗 [JaneGPT-v2 on GitHub](https://github.com/Ravindu-S/JaneGPT-v2)

---

## Created By

**Ravindu Senanayake**

Built from scratch — architecture, tokenizer, and training 
pipeline designed and implemented by the author.

[![GitHub](https://img.shields.io/badge/GitHub-Ravindu--S-black?logo=github)](https://github.com/Ravindu-S)

---

## License

Apache-2.0 (see [LICENSE](LICENSE)).