Hugging Face's logo

Subject-Emu-5259
/
NeuralAI

Text Generation
PEFT
Safetensors
English
llama
lora
conversational
smollm2
dpo
fine-tuned
Model card Files
xet
 Community
NeuralAI / docs /ROADMAP.md
Subject-Emu-5259's picture
Subject-Emu-5259
Push NeuralAI project files - training data, scripts, services, knowledge base
38b4eff verified
preview code
|
Raw
History Blame
6.17 kB

🚀 NeuralAI Development Roadmap

Version Target: 6.0 (The Workstation Pivot) Last Updated: June 6, 2026

✅ Completed Milestones (NeuralAI Legacy)

Phase 0: Core System ✓

  • SmolLM2-360M base model fine-tuned with QLoRA

  • Chat streaming (SSE) working

  • Web UI deployed (NeuralAI → NeuralAI v1.0)

  • Unified Service Migration: Consolidated Model + UI + Terminal into file neural_core_service.py

  • Fixed Chat streaming and Terminal consistency

  • Improved SSE responsiveness with tool execution indicators and non-blocking stream chunks.

Phase 1: Tool Ecosystem ✓

  • Code Execution Sandbox, File Manager, Web Fetcher, DB Connector, Git Assistant

  • Tool detection and routing in chat

Phase 2: DPO Alignment ✓

  • DPO training pipeline implemented

  • Preference dataset expanded to 244 pairs (v12.0)

  • Memphis Culture & Founder Context integration

🏗️ Phase 3: NeuralAI Evolution (In Progress)

1. Workstation Orchestration

  • Establish distinction: NeuralAI (Model) vs. NeuralAI (Hub)

  • UI Overhaul: Added "Workstation Dashboard" tab with project/model/shell status

  • Robust Multi-Turn Context Support (10-message sliding window)

  • Integrated Multi-Modal Speech-to-Speech (Gemini Live + ElevenLabs Fallback)

  • Transition UI from Chat-Only to Multi-Panel Workstation (Expand dashboard features)

  • Implement System-Wide Context Layer

  • Add "Vibe Stack" Workflow Registry

2. Neural Knowledge Graph

  • Implement Persistent Memory (Graph-based)

  • Automate Infrastructure Learning

  • Sync with Supermemory

📊 System Status

  • Main Service: READY (Unified file neural_core_service.py)
  • Voice Service: READY (ElevenLabs v2 Migrated)
  • Model: SmolLM2-360M-Instruct + DPO v14.0 (In Progress)
  • Context: System-wide (Expanding)

Latest DPO Run (v13.0)

  • Training samples: 363
  • Validation samples: 41
  • Epochs: 3
  • Best checkpoint: checkpoint-69
  • Best validation loss: 0.7232
  • Eval progression: 1.6713 → 0.8969 → 0.7232
  • Alignment improvement: 56.73% reduction from first eval to best checkpoint
  • Completed: 2026-05-17 08:41 UTC

🎯 Next Steps (Priority Order)

1. Training Data Expansion

Status: In progress (500+ samples reached, target 1000+)

Categories to expand:

  • Symbolic Logic & Formal Proofs: +50 samples
  • Security & Vulnerability Analysis: +50 samples
  • Multi-Step Algorithmic Reasoning: +50 samples
  • Advanced Mathematics (Calculus/Linear Algebra): +50 samples

2. Evaluation Suite

Status: Created, pending execution

Benchmarks:

  • Code correctness: Generated code runs
  • Response helpfulness: Quality scoring
  • Safety: Refuses harmful requests
  • Latency: Inference speed

🚀 Future Phases (The Agentic Horizon)

Phase 4: Agentic Autonomy & Computer Use

Goal: Transition from "Assistant" to "Operator"

  • Browser Agent Integration: Implement autonomous web navigation and interaction (Computer Use).

  • Multi-Agent Orchestration: Ability to spawn and manage specialized sub-agents for parallel task execution.

  • Long-Horizon Planning: Implement hierarchical planning for tasks requiring 10+ steps.

  • Third-Party App Integration: Direct agentic control over productivity tools (Calendar, Email, CRM).

Phase 5: Universal Knowledge Integration (The "World-Brain" Training)

Goal: Massive expansion of general-world intelligence and cultural context.

  • Natural World: Plants, animals, creatures, ecosystems, and biology.

  • Humanity & Culture: History, religions, beliefs, sociology, and anthropology.

  • The Arts: Music theory, cinematic history, fine arts, and literature.

  • Global Systems: Geography, geopolitics, economics, and planetary sciences.

Phase 6: Model Capability Upgrades

Goal: Integration of frontier reasoning and multimodal capabilities.

  • Deep Reasoning Integration: Implement "Think" modes for complex mathematical and logical deduction.

  • Native Multimodal Understanding: Unified processing of video, audio, and images in a single context window.

  • Test-Time Compute Optimization: Optimize inference to allow the model to "think longer" for harder problems.

📊 Data Files 

data/
├── train.jsonl              # 347 original samples
├── train_v3.jsonl           # 404 samples (latest)
├── train_dpo.jsonl          # 13 DPO pairs
├── train_dpo_expanded.jsonl # 31 DPO pairs
└── train_expanded.jsonl     # 363 samples

📁 Project Structure 

NeuralAI/
├── checkpoints/final_model/    # LoRA adapter
├── data/                       # Training data
├── eval/benchmarks.py          # Evaluation suite
├── from-scratch/web_ui/        # Flask app + static files
│   ├── app.py                  # Main Flask server
│   ├── neuralai_engine.py      # Model + tools
│   └── neuralai_router.py      # Routing logic
├── tools/                      # Tool implementations
│   ├── code_sandbox.py
│   ├── file_manager.py
│   ├── web_fetcher.py
│   ├── db_connector.py
│   └── git_assistant.py
└── training/                   # Training scripts
    ├── train_dpo.py
    ├── generate_training_v3.py
    └── NeuralAI_TPU_Training.ipynb

🔗 Quick Links

📝 Commands 

# Start the service
cd /home/workspace/Projects/NeuralAI/from-scratch/web_ui
python3 app.py

# Generate v5 DPO data
python3 training/generate_dpo_v5.py

# DPO training (currently running in background)
python3 training/train_dpo.py

Next Session Goals:

  1. Run evaluation benchmarks
  2. Expand training data to 1000+ samples
  3. Request GPU or prepare Colab notebook for DPO training