RavinduSen/JaneGPT-v2

JaneGPT v2 — Intent Classification Model

A lightweight, fast, and accurate intent classification model built from scratch for virtual assistant command understanding.

7.8M parameters | 22 intent classes | 88.6% validation accuracy | ~50ms inference on GPU

---

Why I Built This

I'm building JANE — a fully offline, privacy-first AI voice assistant. Llama 3 8B was causing 10–22 second delays for simple commands like "turn up the volume."

That's not a voice assistant. That's a waiting game.

So I designed JaneGPT v2 from scratch — a model that does exactly one job, does it fast, and runs on consumer hardware without any cloud dependency.

---

Model Details

Property	Value
Architecture	Decoder-only Transformer + Classification Head
Parameters	~7.8M
Embedding dim	256
Attention heads	8
KV heads (GQA)	4
Layers	8
FF hidden dim	672
Max sequence length	256
Vocab size	8,192
Tokenizer	Custom BPE
Training accuracy	~96.7%
Validation accuracy	88.6%
Checkpoint size	~30MB

---

Architecture Decisions & Why

Choice	Reason
GQA (4 KV heads, 8 attention heads)	Reduces memory without losing expressiveness
RoPE positional encoding	Better length generalization than learned embeddings
SwiGLU activation	Smoother gradients than ReLU at this model size
RMSNorm	Simpler and faster than LayerNorm
Custom BPE tokenizer	Trained specifically on command-style text

---

Supported Intents (22 classes)

Category	Intents
Volume	volume_up, volume_down, volume_set, volume_mute
Brightness	brightness_up, brightness_down, brightness_set
Media	media_play, media_pause, media_next, media_previous
Apps	app_launch, app_close, app_switch
Browser	browser_search
Productivity	set_reminder, screenshot
Screen	read_screen, explain_screen
Control	undo, quit_jane
Conversation	chat

---

Performance

Input	Predicted Intent	Confidence
"increase the volume"	volume_up	86%
"make it louder"	volume_up	90%
"turn down the brightness"	brightness_down	80%
"open chrome"	app_launch	98%
"play some music"	media_play	96%
"search for cats on youtube"	browser_search	94%
"set a reminder for 5 minutes"	set_reminder	96%
"take a screenshot"	screenshot	88%
"undo that"	undo	92%
"hello"	chat	97%

---

Quick Start

Installation

git clone https://huggingface.co/RavinduSen/JaneGPT-v2
cd JaneGPT-v2
pip install -r requirements.txt

Basic Usage

from classifier import JaneGPTClassifier

classifier = JaneGPTClassifier()

intent, confidence = classifier.predict("turn up the volume")
print(f"Intent: {intent}, Confidence: {confidence:.2%}")
# Output: Intent: volume_up, Confidence: 86.10%

intent, confidence = classifier.predict("open chrome")
print(f"Intent: {intent}, Confidence: {confidence:.2%}")
# Output: Intent: app_launch, Confidence: 98.10%

With Conversation Context

intent, confidence = classifier.predict(
    "not enough",
    context={"last_intent": "volume_up"}
)
# Output: Intent: volume_up, Confidence: 79.00%

---

Training Setup

Component	Details
Hardware	NVIDIA RTX 3050Ti (4GB VRAM)
CPU	AMD Ryzen 9 5900HX
RAM	16GB
Additional	Google Colab (extended training runs)
Framework	PyTorch 2.0+
Training data	Custom command dataset (claude assisted generation under author supervision)

---

Limitations

• Intent classification only — does not generate text
• 22 classes — commands outside supported set classified as chat
• English only
• Optimized for short inputs (1–15 words)
• No entity extraction — returns intent label only

---

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

This model is the intelligence core of JANE — a fully offline, privacy-first AI voice assistant.

🔗 JANE AI Assistant on GitHub 🔗 JaneGPT-v2 on GitHub

---

Created By

Ravindu Senanayake — Computer Science Undergraduate, Sri Lanka

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