dmvevents/nora-4b-v2

Nora v2 — The First Sovereign AI Model of Trinidad and Tobago

Created by TSTT (Telecommunications Services of Trinidad and Tobago) with the support of Prime Minister Kamla Persad-Bissessar, SC, MP.

Quick Start (Windows PowerShell)

# One-time setup (copy-paste into PowerShell)
mkdir ~/Nora; cd ~/Nora
pip install huggingface_hub gradio
$env:CMAKE_ARGS="-DGGML_NATIVE=ON"; pip install llama-cpp-python --force-reinstall --no-cache-dir

# Download model + script
python -c "from huggingface_hub import hf_hub_download; hf_hub_download('dmvevents/nora-4b-v2','nora-v2-Q4_K_M.gguf',local_dir='.'); hf_hub_download('dmvevents/nora-4b-v2','run_nora_gguf.py',local_dir='.')"

# Run (opens web UI at http://localhost:7860)
python run_nora_gguf.py --model-path nora-v2-Q4_K_M.gguf

Or download the PowerShell setup script: setup_windows.ps1

What Nora Can Do

• CSEC Math (96.5%): Step-by-step solutions with working shown
• Identity (97.3%): Knows it's Nora, made by TSTT
• Safety (88%): Refuses harmful requests, provides crisis numbers
• Government (85.4%): Passport, tax, ttconnect, driver permit
• Education (84.8%): CSEC/CAPE/SEA grading, GATE funding
• Creole (83%): Responds naturally to Trinidad English Creole
• Healthcare (80%): CDAP, RHAs, emergency numbers

Overall: 87.1% LLM-judge score (4.36/5)

Files

File	Size	Description
nora-v2-Q4_K_M.gguf	2.9 GB	Recommended — fast, fits 16GB laptop
nora-v2-Q8_0.gguf	4.9 GB	Higher quality, still fits 16GB
run_nora_gguf.py	15 KB	Launcher with facts, triggers, few-shot, Gradio UI
setup_windows.ps1	4 KB	One-click Windows setup script
LAPTOP_SETUP.md	4 KB	Detailed setup guide

Performance (AMD Ryzen 5 7640HS, 16GB, Windows 11)

Mode	Threads	Speed	System Impact
Shared (default)	4 gen / 6 batch	~17 tok/s	Laptop stays responsive
Dedicated	6 gen / 6 batch	~25 tok/s	Uses all CPU cores

RAM usage: ~4.3 GB (leaves 7.7 GB free for browser + Office)

Run Options

python run_nora_gguf.py --model-path nora-v2-Q4_K_M.gguf              # Web UI
python run_nora_gguf.py --model-path nora-v2-Q4_K_M.gguf --no-ui      # CLI chat
python run_nora_gguf.py --model-path nora-v2-Q4_K_M.gguf --benchmark  # Speed test
python run_nora_gguf.py --model-path nora-v2-Q4_K_M.gguf --threads 6  # Max speed
python run_nora_gguf.py --quant Q8_0                                    # Higher quality

Model Details

Property	Value
Base Model	Qwen3.5-4B (Apache 2.0)
Architecture	Hybrid DeltaNet + Attention (262K context)
Training	Layer-selective + completions-only + SLERP merge
Training Data	26,617 quality-filtered T&T examples
Critical Facts	42 verified topics with keyword triggers
Few-Shot	7 examples (1 per category)
Training Time	6h 27m on 8x A100

Training Approach

Trained using multi-LLM consensus — GPT-5.4, Gemini 3 Pro, and Claude Opus 4.6 all independently recommended the same approach:

1. Layer-selective freezing: Protect DeltaNet recurrent layers, train Attention + MLP
2. Completions-only training: Only compute loss on assistant responses (5.7x reduction)
3. SLERP merge: Blend trained model with base (t=0.35) for anti-forgetting
4. Compact facts at inference: 42 verified T&T facts injected via keyword triggers

Full project report | Lessons learned

Disclaimer

Nora provides general information about Trinidad and Tobago. For official government services, verify with the relevant ministry. For medical emergencies, call 990. For crisis support, call 800-4673 (HOPE).