Subject-Emu-5259
/

NeuralAI

@@ -1,76 +1,236 @@
 ---
-base_model: HuggingFaceTB/SmolLM2-360M-Instruct
 library_name: peft
-tags:
-- lora
-- dpo
-- smollm2
-- text-generation
-- transformers
-- trl
-- neuralai
 license: apache-2.0
-language:
-- en
-pipeline_tag: text-generation
-model_name: NeuralAI v17-dpo (D17)
-collection: https://huggingface.co/collections/Subject-Emu-5259/neuralai-model-family-6a66ee29c7c5f26e044dee3c
 ---
-# NeuralAI — D17 DPO LoRA Adapter (v17) 🧠
-NeuralAI is a DPO-aligned generative AI engine by **De'Andrew Preston Harris**.
-This repository publishes the **D17 DPO LoRA adapter** (v17), a continuation of
-the v16 adapter, trained on top of `HuggingFaceTB/SmolLM2-360M-Instruct`.
-This is the **main production adapter** in the NeuralAI model family.
-## NeuralAI Model Family
-| Model | Role | HF Repo |
-| --- | --- | --- |
-| **NeuralAI (this adapter)** | Main 360M DPO-aligned LoRA | `Subject-Emu-5259/NeuralAI` |
-| NeuralAI-Air-135M | Custom 135M base language model | `Subject-Emu-5259/NeuralAI-Air-135M` |
-| NeuralAI-Air-135M-SFT | Supervised fine-tune of the 135M base | `Subject-Emu-5259/NeuralAI-Air-135M-SFT` |
-Collection: [NeuralAI Model Family](https://huggingface.co/collections/Subject-Emu-5259/neuralai-model-family-6a66ee29c7c5f26e044dee3c)
-## Training (D17 / v17 — 2026-07-20)
-- **Base model:** `HuggingFaceTB/SmolLM2-360M-Instruct` (360M)
-- **Method:** DPO (TRL), LoRA r=32 / alpha=64, dropout 0.05
-- **Seed adapter:** v16 (`checkpoints/v2_model`)
-- **Dataset:** 679 preference pairs (`data/train_dpo_v16_combined.jsonl`)
-- **Epochs:** 3 — **Steps:** 129 — **Duration:** ~31 min (Colab GPU)
-- **Final train loss:** ~0.39 (0.692 → 0.396)
-- **Reward accuracy:** 0.975 (chosen preferred over rejected)
-- **Reward margin:** ~0.9 (stable, no collapse)
-- **Entropy:** ~2.15–2.33 (stable)
-D17 is a DPO *continuation* of v16 — an alignment refinement, not new knowledge
-injection. Across the 679-pair distribution it sharpened preferences toward
-correct code, safe refusals, identity-accurate responses, sound math/reasoning,
-clean code style, concise answers, and proper tool usage, while down-ranking
-rejected patterns. Chosen-response rewards rose (~0.54 → 0.69) and rejected
-rewards fell (~−0.19 → −0.26) with a healthy ~0.9 margin and stable entropy —
-textbook successful DPO with no sign of reward hacking or mode collapse. No
-held-out eval set was configured, so generalization is inferred from training
-reward signals only.
-## Usage (PEFT)
 ```python
-from peft import PeftModel
-from transformers import AutoModelForCausalLM, AutoTokenizer
-base = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM2-360M-Instruct")
-model = PeftModel.from_pretrained(base, "Subject-Emu-5259/NeuralAI")
-tokenizer = AutoTokenizer.from_pretrained("Subject-Emu-5259/NeuralAI")
 ```
-## Links
-- **GitHub:** https://github.com/Subject-Emu-5259/NeuralAI
-- **Live demo:** https://neuralai-web-ui-deandrewharris.zocomputer.io
-- **Model family collection:** https://huggingface.co/collections/Subject-Emu-5259/neuralai-model-family-6a66ee29c7c5f26e044dee3c
-- **Founder:** De'Andrew Preston Harris (D. Harris / Dre)

 ---
+language:
+  - en
 library_name: peft
 license: apache-2.0
+tags:
+  - lora
+  - conversational
+  - text-generation
+  - peft
+  - fine-tuned
+  - mamba
+  - ssm
+  - neuralai
+  - base-model
+model_id: Subject-Emu-5259/NeuralAI
+inference: false
+---
+# 🧠 NeuralAI: The Generative AI Engine
+<p align="center">
+  <img src="neuralai_banner.svg" alt="NeuralAI — Your AI. On your hardware. In your browser." />
+</p>
+<p align="center">
+  <a href="https://github.com/Subject-Emu-5259/NeuralAI"><img src="https://img.shields.io/badge/GitHub-NeuralAI-181717?logo=github" alt="GitHub"></a>
+  <a href="https://huggingface.co/Subject-Emu-5259/NeuralAI"><img src="https://img.shields.io/badge/🤗%20Hub-NeuralAI-FFD21E" alt="HF Hub"></a>
+  <a href="https://huggingface.co/Subject-Emu-5259/NeuralAI-Mamba-K1"><img src="https://img.shields.io/badge/🤗%20Hub-Mamba%20K1-blue" alt="Mamba K1"></a>
+  <a href="https://huggingface.co/Subject-Emu-5259/NeuralAI-Mamba-K2"><img src="https://img.shields.io/badge/🤗%20Hub-Mamba%20K2-green" alt="Mamba K2"></a>
+  <a href="https://neuralai-web-ui-deandrewharris.zocomputer.io"><img src="https://img.shields.io/badge/Live-NeuralAI%20Web%20UI-6366f1" alt="Live UI"></a>
+</p>
+---
+## 📊 Repository Composition
+| Language | Percentage |
+| --- | --- |
+| Python | 71.1% |
+| HTML | 13.0% |
+| JavaScript | 12.4% |
+| CSS | 2.6% |
+| Shell | 0.4% |
+| Jupyter Notebook | 0.3% |
+| Jinja | 0.2% |
+**The High-Velocity AI Engine for Your Entire Vibe Stack**
+NeuralAI is the central intelligence engine developed by **De'Andrew Preston Harris**. Conceived and engineered as an owned AI platform, it spans fine-tuned transformer models, custom SSM base models, DPO alignment, and a production web UI — all designed for local-first, private AI computing.
+---
+## 🏗️ Model Family
+```mermaid
+graph TB
+    subgraph "NeuralAI Model Family"
+        direction TB
+        K1["🧬 Mamba K1<br/>130M · SFT 50 steps<br/>First Owned Base"]
+        K2["🧬 Mamba K2<br/>793M · Q4_K_M GGUF<br/>460MB · Production"]
+        K3["🔬 Mamba K3<br/>SFT 500-1000 steps<br/>10K+ UltraChat<br/>In Training"]
+        K1 --> K2 --> K3
+        K2 --> PROD["🚀 Active Inference<br/>LM Studio · 460MB RAM<br/>neuralai-web-ui.zocomputer.io"]
+    end
+    style K1 fill:#4a90d9,color:#fff
+    style K2 fill:#22c55e,color:#fff
+    style K3 fill:#f59e0b,color:#000
+    style PROD fill:#10b981,color:#fff
+```
+### Complete Lineup
+| Model | Architecture | Params | Training | Status | Location |
+|-------|-------------|--------|----------|--------|----------|
+| **Mamba K1** | Mamba SSM | 130M | SFT LoRA 500 steps on 1K UltraChat (intel format) | 🔄 Retraining for chat coherence | `Subject-Emu-5259/NeuralAI-Mamba-K1` |
+| **Mamba K2** | Mamba SSM | 793M | Base pretrained — SFT queued (Q4_K_M GGUF) | ⚠️ Base model only | `Subject-Emu-5259/NeuralAI-Mamba-K2` |
+| **Mamba K3** | Mamba SSM | 2.8B | Base pretrained — SFT queued | ⚠️ Base model only | local `models/mamba-k3-base/` |
+### Why Mamba SSM
+| Property | Benefit |
+|----------|--------|
+| Complexity | \(O(n)\) linear — scales to long context efficiently |
+| Inference | Fast at any sequence length, not just short prompts |
+| Memory | K2 runs at 460MB (Q4_K_M GGUF) — fits on any device |
+| Ownership | NeuralAI trains and merges all release weights on top of public Mamba SSM bases — every released GGUF is a fully merged model, not a raw base |
+| Ecosystem | LM Studio, llama.cpp, Hugging Face — mature deployment options |
+---
+## 🌟 Vision & Manifesto
+NeuralAI doesn't just predict text; it *operates the work*. The core mission is to create a multimodal generative system that bridges the gap between raw idea and execution. By fusing autoregressive generation with adaptive agency, NeuralAI becomes more than a chatbot — it is a persistent, reasoning partner.
+Born from resilience and ambition in Memphis, Tennessee and West Memphis, Arkansas, NeuralAI represents a forward-thinking approach to personal, private AI computing.
+---
+## 🛠️ Tech Stack & Architecture
+NeuralAI is built on a high-performance architecture that decouples the inference engine from the web interface, enabling lightweight cloud hosting with powerful local inference.
+### Core Stack
+- **Production Model**: Mamba K2 793M Q4_K_M GGUF (460MB) via llama.cpp — active inference engine
+- **Model Family**: Mamba K1 (130M, SFT LoRA retry) → Mamba K2 (793M base GGUF) → Mamba K3 (2.8B base)
+- **Inference Engine**: llama.cpp server with a custom `neuralai-intel` chat format (vocabulary-friendly for GPT-NeoX / Mamba tokenizers)
+- **Vocal Identity**: Andrew (Warm/Multilingual) — Optional voice synthesis integration
+- **Web Interface**: Custom Flask UI served via Zo Computer at `neuralai-web-ui-deandrewharris.zocomputer.io`
+- **Tool Chain**: 10 live slash commands (/web, /fetch, /browse, /research, /img, /speak, /summarize, /translate, /news, /yt) + NL→Tool Router
+### Future Scale Path
+| Stage | Params | Goal | Status |
+|-------|--------|------|--------|
+| **Mamba K1 SFT v2** | 130M | Finish 500-step SFT with intel format + GGUF | 🔄 Active |
+| **Mamba K2 SFT** | 793M | SFT LoRA 500 steps → merge → Q4_K_M GGUF | 📋 Next |
+| **Mamba K3 SFT** | 2.8B | SFT LoRA 1000 steps → merge → Q4_K_M GGUF | 📋 Next |
+| **Mamba 2B/3B** | ~2-3B | Scaled SSM architecture, benchmarks | 📋 Planned |
+---
+## ✨ Key Features & Capabilities
+### 💬 Multimodal Chat & Agentic Intelligence
+- **High-Velocity Text Inference**: Fast, local inference with deep context awareness
+- **Deep Reasoning Mode**: Integration of test-time compute and chain-of-thought reasoning
+- **Autonomous Agentic Workflows**: Agent-mode interaction with browser, terminal, and third-party apps
+- **Live S2S (Speech-to-Speech)**: Real-time voice interaction with integrated microphone interface
+- **Identity Vault & Memory**: Persistent user memory and rule constraints
+### 🔧 Developer & Engineering Tools
+- **10 Web Tool Commands**: Search, fetch, browse, research, image gen, TTS, summarize, translate, news, YouTube
+- **NL→Tool Router**: Natural language web requests auto-routed to the right tool
+- **Model Manager**: CLI switching between all registered models
+- **Benchmark Suite**: Perplexity, generation diversity, MMLU-style, reasoning tests
 ---
+## 🚀 Model Lineage
+```mermaid
+timeline
+    title NeuralAI Model Evolution
+           : Custom 135M base SFT
+    2026 Q3 : Mamba K1 — First owned base
+           : 130M SSM · Proof of Concept
+    2026 Q3 : Mamba K2 — Scaled base
+           : 790M Q4_K_M · GGUF ready
+    2026 Q3 : Mamba K3 — Full SFT
+           : 500-1000 steps · 10K+ samples
+    2026 Q4 : Mamba 2B/3B — Next scale targets
+           : 2B SSM → 3B Core Intelligence
+```
+---
+## 🚀 Deployment
+```bash
+# 1. Start the inference service
+cd NeuralAI
+supervisorctl -c /etc/zo/supervisord-user.conf restart neuralai-lmstudio
+# 2. Start the web UI service
+python3 services/webui_service.py
+```
+### Mamba K2 (LM Studio / llama.cpp)
+```bash
+# Download from HuggingFace
+huggingface-cli download Subject-Emu-5259/NeuralAI-Mamba-K2 \
+  mamba-790m-hf.Q4_K_M.gguf --local-dir ./models/
+# Serve with the NeuralAI chat format
+python3 services/lmstudio_server.py \
+  --model models/mamba-790m-hf.Q4_K_M.gguf \
+  --chat_format neuralai-intel \
+  --port 1234
+```
+### Mamba K1 (Python)
 ```python
+from transformers import MambaForCausalLM, AutoTokenizer
+model = MambaForCausalLM.from_pretrained("Subject-Emu-5259/NeuralAI-Mamba-K1")
+tokenizer = AutoTokenizer.from_pretrained("Subject-Emu-5259/NeuralAI-Mamba-K1")
 ```
+### Containerized Deployments
+| Deployment | Dockerfile | Stack | Status |
+| --- | --- | --- | --- |
+| **Web Chat** | `services/start_lmstudio.sh` + Flask UI | llama.cpp + `neuralai-web-ui.zocomputer.io` | ✅ Live |
+---
+## 🌌 NeuralAI Ecosystem
+The standalone software implementation of the NeuralAI core is **NeuralLabs**:
+👉 [github.com/Subject-Emu-5259/NeuralLabs](https://github.com/Subject-Emu-5259/NeuralLabs)
+**Software Downloads**: Latest beta builds available at:
+👉 [zo.pub/deandrewharris/neurallabs-beta](https://zo.pub/deandrewharris/neurallabs-beta)
+---
+## 📈 Current State & Active Goals
+- **Legacy DPO v17 / Air 135M / SmolLM2-360M**: Retired and removed from the repository
+- **Mamba K1**: First owned SSM base model — retraining with intel-format SFT for coherent chat (Colab/GPU)
+- **Mamba K2**: Base pretrained GGUF ready — awaiting SFT
+- **Mamba K3**: 2.8B base downloaded — awaiting SFT
+- **Last Maintenance**: August 1, 2026 (Mamba Era — fix chat format + retrain pipeline)
+---
+## 👤 Creator
+Built by **De'Andrew Preston Harris** ([@deandrewharris94](https://linkedin.com/in/deandrewharris94/)) with Google Gemini AI Studio/Colab collaboration.
+From Memphis, Tennessee. Raised in West Memphis, Arkansas. AI Software Engineering at Maestro College.
+---
+# NeuralAI → Hugging Face sync is live