๐Ÿง  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."

License: MIT Base Model: Gemma-2-9B Training: QLoRA Unsloth Format: GGUF Demo: Gradio


๐ŸŽฏ What does it solve? (The Real Problem)

Standard AIs (GPT, Claude, Gemini, Llama-base) fail miserably with neurodivergent minds:

  • ๐Ÿ—ฃ๏ธ Neurotypical Language: Performative "empathy", generic advice ("make a list", "take a deep breath"), structured social moralizing.
  • ๐Ÿงฉ ND Context Blindness: Zero understanding of Task Paralysis, RSD (Rejection Sensitive Dysphoria), Alexithymia, Time Blindness, Autistic Burnout, Masking, and 2e/3e profiles (Twice-Exceptional).
  • ๐Ÿ“‰ Lack of Logical Structure: They give you "tips", not Mathematical Solution Paths or action algorithms.

โš™๏ธ The Solution: Psychological Mathematical Resolution Principle (PMR)

Nehuronion doesn't "answer". It calculates.

  1. Input: Your cognitive chaos or saturation (Text, raw idea flows, paralysis).
  2. 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.
  3. 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.

๐Ÿ—ฃ๏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. ๐Ÿง โšก"


๐Ÿ—๏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

๐Ÿš€ Quick Start (Zero Setup)

1. ๐Ÿค— Gradio Demo (Browser - Free)

๐Ÿ‘‰ OPEN INTERACTIVE DEMO ON HF SPACES

2. ๐Ÿ’ป 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. ๐Ÿ’ป 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"

๐Ÿงช 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 โž” Timer 5' โž” 1 isolated micro-task. Deploy running. ๐ŸŽฎ"
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. ๐Ÿงฉ"
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. ๐Ÿง "

๐Ÿ“œ 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.

๐Ÿค Collaborate (Goal: Open ND Ecosystem)

  1. โญ Star the Repo on Hugging Face to boost its visibility in the global feed.
  2. ๐Ÿ› Issues: Report neurotypical tone deviations, mathematical hallucinations, or uncovered cognitive profiles.
  3. ๐Ÿงฌ Format Optimization: Help port the cognitive matrices to optimized environments like MLX (Apple Silicon) and EXL2.

Made with ๐Ÿง ๐Ÿ’œ by iangel094 | Developed reading frontier research, optimizing with Unsloth, and without asking traditional corporations for permission.

Downloads last month
25
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for nehuronion-dvt/nehuronion-dvt-9b

Finetuned
(330)
this model