Port 5000 Operating Environment

Unified Web Model Builder & Live Visualizer Suite

ImpressionCore adheres to Directive 4: Strict Architectural Separation. The Model Builder operates as an independent, high-performance Flask & WebGL environment on Port 5000, orchestrating the entire 10-step AI lifecycle before export to the Port 8000 production runtime.

ImpressionCore Unified Web Model Builder Suite

Glassmorphism Interactive Builder: Preset Auto-Population, 3D Node Orbit, and Real-Time Loss Monitor FIGURE 4.0 • BUILDER SUITE

Lifecycle Engineering

The 10-Step Training & Distillation Pipeline

A comprehensive, verifiable end-to-end workflow taking raw data from preflight verification to quantized edge deployment.

Step 01 • Hardware Verification

GPU Preflight & Capability Audit

Automatically interrogates system hardware: CUDA capability, compute architecture (Pascal, Ampere, Ada Lovelace), VRAM headroom, and PyTorch environment health. Guarantees operations strictly adhere to the <4GB memory budget.

Step 02 • Data Pipeline

Streaming Ingestion & Validation

Loads multi-format training sets (JSONL, raw text corpora, audio features, visual embedding grids) with streaming chunk validation and SHA-256 dataset hashing.

Builder Live Preflight and Pipeline Controller

Step 1 & 2 Live Controller: GPU Preflight & Pipeline Step Selector UI STEP 1-2

Builder Live Model Definition and Presets

Step 4 Architecture Config: Canonical Presets with Live VRAM Budget Bar UI STEP 3-4

Step 03 • Tokenization

Domain-Adapted BPE Tokenizer

Configures custom Byte-Pair Encoding tokenizers with domain-adapted vocabulary tables (50,257 tokens), special token injection, and sub-word compression profiling.

Step 04 • Architecture Config

Automated Preset Population

Select any canonical tier (`B1 Hope`, `B2 Insight`, `B3 Apex`, `B3 Ultra`) to instantly populate hidden layers, dimension size, attention heads, and optimizer parameters with live hardware headroom bars.

Step 05 • Knowledge Distillation

Teacher-Student Knowledge Distillation

Distills heavy teacher models (DialoGPT-medium, Qwen2.5) into compact student weights via Kullback-Leibler (KL) divergence and feature representation matching.

Step 06 • Loss & Annealing

Cosine Annealing & Loss Telemetry

Real-time monitoring of AdamW optimization, Cosine Annealing learning rate schedules, gradient norm clipping, and loss convergence curves.

Builder Live Training Progress and Telemetry

Step 6 Training Telemetry: Live Loss Curves, Step Counter, and GPU VRAM Meter UI STEP 5-6

Step 07

Checkpoint Browser

Automated background checkpointing with SHA-256 cryptographic integrity hashes for non-repudiation logging.

Step 08

Evaluation Audit

Validation against the 10/10 conversational benchmark, perplexity testing, and epistemic uncertainty checks.

Step 09

GGUF Quantization

One-click export to quantized formats (`Q4_K_M`, `Q8_0`, `FP16`) for instant llama.cpp edge execution.

Step 10

Production Serving

Zero-copy model handoff to the FastAPI/Vite runtime on Port 8000 with real-time health telemetry.

Developer Tooling Suite

Deep Model Introspection & Real-Time Inspection

The Port 5000 builder provides unmatched visibility into neural activations, tensor shapes, and underlying PyTorch source code.

Tensor Shape Tracer: Verifying Forward Pass Dimensions & Latent Projections SHAPE TRACER

Tensor Shape Tracer

Traces multi-dimensional tensor activations across every attention head, latent projection, and FFN layer to eliminate shape mismatches before initiating long training runs.

PyTorch Code Mapper: Correlating Visual UI Parameters with Clean Python Source Code CODE MAPPER

Live PyTorch Code Mapper

Dynamically translates selected visual layers, attention parameters, and optimizer settings into production-ready PyTorch modules following the Clean Code directive.

Live Model Assistant: Interactive Socratic Dialogue & Hyperparameter Guidance BUILDER CHAT

Live Model Assistant

Embedded conversational assistant for querying training loss patterns, tuning hyperparameters, and debugging architecture bottlenecks interactively.

ImpressionCore 3D Architecture Node Orbit

3D Neural Architecture Orbit: Visualizing Forward Pass Connectivity 3D NODE ORBIT

3D Neural Architecture Orbit

Interactive 3D WebGL node graph visualizing layer depth, attention connections, and tensor routing across the neural manifold.

Quickstart

Launching the Builder Studio Locally

Get the full Port 5000 Unified Web Builder running on your local machine with automated preflight checks in seconds.

Terminal / PowerShell Execution Windows & Linux

# 1. Launch System A: Unified Model Builder Studio (Port 5000)
launch_builder.bat

# Or via Python directly:
python src/interfaces/web/server.py --port 5000

# 2. Automated End-to-End Builder Verification (All 9 functions tested)
python src/dev_tools/exercise_builder_site.py

Explore Canonical Models → View Documentation Hub