Announcing MoroAI v0.1.0•Open Source Local Model Foundry

Build Private AI. Own Your Models.

MoroAI transforms private datasets into dependable, high-yield local language models. Equipped with information-theoretic data compilation, VRAM pre-flight prediction, and automatic 5-step self-healing OOM recovery.

$pip install moroai
Unified CLI + Mission Control
Self-Healing OOM Recovery
Cryptographic Provenance
Closed-Loop DPO Flywheel
$moro train --config moro.yaml
[INIT] Moro Engine v0.1.0 starting training loop with Auto-Recovery daemon...
Epoch 1/3 | Step 100/1200 | Loss: 1.842 | Perplexity: 6.31 | VRAM: 7.4GB
[OOM-GUARD TRIGGERED] Spurious CUDA spike detected at step 340 (seq_len burst)
⚡ Executing Step 2 Recovery: Emptied CUDA cache + reduced micro-batch to 1 + increased accum to 16
[RESUME] Training resumed seamlessly from step 339 checkpoint without crash!
Epoch 3/3 | Step 1200/1200 | Loss: 0.741 | Convergence verified
✓ Training completed successfully in 28m 14s. Final adapter: ./checkpoints/run_01/
$█
Subsystems & Capabilities

The Complete Lifecycle for Private AI Models

Other tools only provide a training script. MoroAI is a complete, reliable, production-grade local adaptation pipeline from raw enterprise documents to audited deployment.

Core Innovation

Epistemic Data Compiler

Information-theoretic curation with MI Guard. Rare boundary facts and domain specifics are mathematically preserved rather than pruned as outliers.

Explore deep dive
Zero Crash

Hardware-Aware Recipes

Pre-flight VRAM audit predicts memory requirements down to the megabyte. Automatically configures LoRA rank, quantization, and batching for your exact GPU.

Explore deep dive
Autonomous

Self-Healing Training

5-step automatic OOM recovery engine. When memory spikes or loss explodes, MoroAI empties cache, resizes gradients, and rolls back checkpoints automatically.

Explore deep dive
Safety & Quality

Multi-Layered Evaluation

Deterministic validation rules, semantic accuracy judges, Wasserstein drift detection, and adversarial perturbation stress tests prove model stability.

Explore deep dive
Continuous

DPO Feedback Flywheel

Turn production queries and corrections into preference pairs. Close the learning loop locally so your model continuously improves with zero cloud exposure.

Explore deep dive
Instant Serving

One-Click Deployment

Directly export to GGUF format with customized quantization (Q4_K_M, Q8_0) and register into local Ollama, vLLM, or Docker in a single CLI command.

Explore deep dive
Audit Ready

Cryptographic Provenance

Full Software Bill of Materials (SBOM) with cryptographic hashes linking the compiled dataset, hyperparameter recipe, and training weights for audit compliance.

Explore deep dive
Interactive UI

Mission Control Dashboard

Modern, reactive web control center. Drag-and-drop training datasets, watch real-time GPU metrics and loss curves, and trigger evaluations visually.

Explore deep dive
Insights

Experiment Analytics

Cross-run comparative analytics and hyperparameter Pareto frontier visualization. Discover which data slices yielded the highest benchmark jump.

Explore deep dive
Extensible

Modular Plugin System

Clean Python entry-points for custom tokenizers, data loaders, loss criteria, and enterprise deployment backends. Extend without modifying core code.

Explore deep dive
Consumer Hardware

Optimized for Consumer GPUs

Built specifically to run 1.5B to 14B parameter models on single consumer graphics cards (RTX 3060, 4070, Apple Silicon M-series, or AMD ROCm).

Explore deep dive
Absolute Privacy

100% Private & Sovereign

True local-first execution. Training data, model weights, and telemetry never leave your workstation or private VPC. No external API dependencies.

Explore deep dive
End-to-End Pipeline

The MoroAI Architecture

Click on any phase below to inspect the mathematical guarantees, CLI commands, and contracts linking raw enterprise data to self-improving local models.

STAGE 01

Epistemic Compiler

Compiles raw logs, code, or knowledge bases into high-density training pairs. Protects rare domain edge cases with Mutual Information Guard.

Input:Raw JSONL, Markdown, CSV, DB exports
Output:curated_train.jsonl with quality telemetry
+28% Info Entropy
0 Rare Facts Lost
3.4x Token Density
Execution Interface
Ready
$moro data build --source ./raw_data.jsonl --mi-guard-threshold 0.85
Subsystem Contracts

Zero external telemetry is sent. The entire process runs under local Python process management with deterministic seeding and file locking.

Read Full Documentation
Competitive Matrix

Why Teams Choose MoroAI

Other tools are mere training wrappers. MoroAI is a complete, self-healing, auditable adaptation foundry engineered for dependable local production.

Capability
MoroAI
Full Platform
AxolotlUnslothLLaMA-Factory
Data Quality & Curation
Epistemic + MI GuardBasic filteringNoneBasic regex
Edge-Case Preservation
Guaranteed (MI Guard)NoNoNo
VRAM Pre-Flight Simulation
Exact Tensor SimulationManual GuessworkVRAM Rule of ThumbNone
Autonomous OOM Auto-Recovery
5-Step Autonomous HealingCrashes on OOMCrashes on OOMCrashes on OOM
Loss-Spike Anomaly Rollback
Automated Checkpoint RevertManualManualManual
Multi-Layer Evaluation
Rules + Drift + LLM JudgeLoss metric onlyLoss metric onlyBasic BLEU/ROUGE
Adversarial Stress Testing
Built-in Perturbation EngineNoNoNo
Cryptographic Provenance (SBOM)
SHA-256 Manifest + Gate AuditNoNoNo
Interactive Mission Control UI
Real-time Web DashboardCLI onlyCLI / NotebookGradio demo
One-Click Local Deployment
Direct Ollama / GGUF ExportManual conversion scriptManual llama.cpp stepsManual export commands
Continuous DPO Flywheel
Closed-Loop Feedback MiningSeparate DPO scriptManual dataset preparationManual pairs
Plugin Architecture
Python setuptools entry_pointsHardcoded scriptsProprietary coreForks required
100%
Local Execution
Zero bytes leave your host
0
Unrecovered OOMs
Guaranteed 5-step mitigation
414
Integration Tests
End-to-end verified suite
< 30m
Consumer GPU Train
1.5B models on 8GB VRAM
Proven in Production

Trusted by Builders of Sovereign AI

Engineered for practitioners who require absolute data privacy and mathematical reliability.

"We spent 3 months dealing with random CUDA OOM crashes using custom scripts. MoroAI predicted our VRAM down to 100MB and the 5-step auto-recovery saved 6 long runs without any human intervention."

Dr. Elena Rostova
Principal ML Engineer
FinTech Sovereign Labs

"MI Guard in the Epistemic Data Compiler is revolutionary. When we pruned our 40,000 internal legal support cases with other tools, rare compliance edge cases vanished. MoroAI mathematically kept them safe."

Marcus Vance
VP of AI Architecture
Lexis Sovereign AI

"Running local adaptation on an RTX 4090 and deploying straight to Ollama in one CLI command transformed our internal prototyping cycle from weeks to under two hours."

Siddharth Patel
Staff Platform Engineer
Autonomous Health
Join The Sovereign AI Movement

Ready to Take Control of Your AI?

Get the latest releases, whitepapers on information-theoretic data curation, and benchmarks for consumer GPU local adaptation directly in your inbox.