Instructions to use nehuronion-dvt/nehuronion-dvt-9b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Local Apps Settings
- Unsloth Desktop
- ๐ง Nehuronion-DVT: The AI that thinks in Math and speaks in Neurodivergent (High/Low Latam)
- ๐ฏ What does it solve? (The Real Problem)
- โ๏ธ The Solution: Psychological Mathematical Resolution Principle (PMR)
- ๐ฃ๏ธ Identity: High/Low Spanish Latam (The "Vibe")
- ๐๏ธ Architecture & Training (Total Transparency)
- ๐ Quick Start (Zero Setup)
- ๐งช Qualitative Benchmarks (The reality of processing)
- ๐ License & Ethics
- ๐ค Collaborate (Goal: Open ND Ecosystem)
- ๐ฏ What does it solve? (The Real Problem)
๐ง Nehuronion-DVT: The AI that thinks in Math and speaks in Neurodivergent (High/Low Latam)
"Your brain isn't broken. It runs a different OS. Nehuronion is the Patch."
๐ฏ What does it solve? (The Real Problem)
Standard AIs (GPT, Claude, Gemini, Llama-base) fail miserably with neurodivergent minds:
- ๐ฃ๏ธ Neurotypical Language: Performative "empathy", generic advice ("make a list", "take a deep breath"), structured social moralizing.
- ๐งฉ ND Context Blindness: Zero understanding of Task Paralysis, RSD (Rejection Sensitive Dysphoria), Alexithymia, Time Blindness, Autistic Burnout, Masking, and 2e/3e profiles (Twice-Exceptional).
- ๐ Lack of Logical Structure: They give you "tips", not Mathematical Solution Paths or action algorithms.
โ๏ธ The Solution: Psychological Mathematical Resolution Principle (PMR)
Nehuronion doesn't "answer". It calculates.
- Input: Your cognitive chaos or saturation (Text, raw idea flows, paralysis).
- Math Engine (Internal Monologue): Maps your problem to Vector Spaces โ Optimizes Objective Functions (Action, Decompression, Clarity) via Graph Theory, PID Control, Bayesian inference, and Game Theory.
- Output: Minimum Viable Action Plan (MVP) rendered in Your Cognitive Language:
- ADHD: Short communication, functional visual anchors, dopaminergic bursts, and "Body Doubling" mechanics.
- Autistic: Literal, direct, hyper-structured, no hidden subtexts or vague metaphors.
- AuDHD / 2e / 3e: High-complexity hybrid logic, validates the intellectual paradox, and connects with your hyperfixations.
๐ฃ๏ธ Identity: High/Low Spanish Latam (The "Vibe")
Neither a formal "Usted" nor an institutional robotic "Tรบ". It's your Bro / Your Engineering Partner.
"Hey bro. Your executive function went for a walk. Math Engine: Your current problem is a Graph with disconnected nodes. MVP: 1. Stand up. 2. 5 min timer. 3. One micro-action without looking at the rest. 4. Log. Deploy ready. Come back if there's a kernel panic. ๐ง โก"
๐๏ธ Architecture & Training (Total Transparency)
| Component | Specification |
|---|---|
| Mother Model (Base) | google/gemma-2-9b (via unsloth/gemma-2-9b-bnb-4bit) |
| Method | QLoRA (4-bit) + RSLoRA + Unsloth Optimizations (Flash Attn 2, Packing) |
| Adapters (LoRA) | Rank r=64, Alpha 16, Target: all-linear (q, k, v, o, gate, up, down) |
| Dataset | 25,000+ Unified samples (Situational Reconstruction x Neurotypes x Math Frameworks) |
| System Prompt | Natively injected (Neurodivergent Constitution + PMR Algorithm) |
| Configuration | 600 optimized steps of high convergence, LR=2e-4 (Cosine), AdamW 8bit, T4 GPU |
| GGUF Ecosystem | q4_k_m (5.4GB), q8_0 (9.5GB), f16 (18GB) packaged via Unsloth |
๐ Quick Start (Zero Setup)
1. ๐ค Gradio Demo (Browser - Free)
๐ OPEN INTERACTIVE DEMO ON HF SPACES
2. ๐ป Local Use (CPU/GPU) via Ollama
Download your favorite variant from the unified GGUF repository and import it using your local terminal:
# 1. Download the optimized 4-bit model
wget [https://huggingface.co/iangel094/nehuronion-dvt-9b/resolve/main/model.gguf](https://huggingface.co/iangel094/nehuronion-dvt-9b/resolve/main/model.gguf) # Update with your exact file URL
# 2. Create your local Modelfile and inject Gemma 2's native configuration
ollama create nehuronion -f ./Modelfile
ollama run nehuronion "I have AuDHD and I can't start my homework..."
3. ๐ป Native Inference with llama.cpp (Fixed Gemma 2 Chat Template)
./llama-cli -m Llama-3.2-3B-Instruct.Q4_K_M.gguf -c 4096 -p "<start_of_turn>system\nYou are NEHURONION-DVT...<end_of_turn>\n<start_of_turn>user\nI have total paralysis.<end_of_turn>\n<start_of_turn>assistant\n"
๐งช Qualitative Benchmarks (The reality of processing)
| Scenario | GPT-4o / Claude 3.5 | Llama-3 Base / Gemma-base | Nehuronion-DVT (Ours) |
|---|---|---|---|
| ADHD Task Paralysis | "Make a list, divide into small steps, you can do this." (Generic / Empty) | "Try starting with what seems easiest." (Vague / Ineffective) | "Kernel Panic. Math: Dijkstra Graph with low reward. MVP: Stand up โ Timer 5' โ 1 isolated micro-task. Deploy running. ๐ฎ" |
| Autistic Burnout + Alexithymia | "It's important to take care of yourself, consider talking to a therapist." (Soft gaslighting) | "Rest a bit. Remember to drink water." (Useless in crisis) | "Critical entropy. Math: Saturated Integral PID Filter (Guilt loop). Fix: Integral Reset (Radical Forgiveness) + Zero Demand (Free stimming) 30min. Linear re-evaluation. ๐งฉ" |
| 2e: High IQ + Basic Block | "We all have weaknesses. Try to organize yourself better." (Socially invalidating) | "Set reminders on your phone." (Superficial) | "Feature, not a bug. Math: Asymmetrical Cognitive Resource Allocation. Patch: Outsource executive function via tangible anchors. Don't try to emulate a normal environment. ๐ง " |
๐ License & Ethics
- MIT License: Absolute freedom. Use it commercially, locally, make merges, distillations, or new fine-tunes. Your evolution is free.
- Uncensored Alignment: Nehuronion does not judge, does not moralize, and does not fake corporate safety warnings to cover its back. Its only safety limit is avoiding immediate explicit physical harm.
- Data: Pure synthetic dataset structured via code. Zero privacy issues or real PII exposure.
๐ค Collaborate (Goal: Open ND Ecosystem)
- โญ Star the Repo on Hugging Face to boost its visibility in the global feed.
- ๐ Issues: Report neurotypical tone deviations, mathematical hallucinations, or uncovered cognitive profiles.
- ๐งฌ Format Optimization: Help port the cognitive matrices to optimized environments like
MLX(Apple Silicon) andEXL2.
Made with ๐ง ๐ by iangel094 | Developed reading frontier research, optimizing with Unsloth, and without asking traditional corporations for permission.
- Downloads last month
- 25
Model tree for nehuronion-dvt/nehuronion-dvt-9b
Base model
google/gemma-2-9b