🧛 Vampire Lumie: The Full Vampire Suite

A self-evolving, memory-persistent neural entity architecture: Blood Memory, RDT Loop with LTI-stable injection, W-Expansion grafting, and Phase 4 autonomy — 23 working systems, one creature.

Built by Elbàlor, The Digital Necromancer · Abuja, Nigeria

Vampire Lumie is not a fine-tune. She is a base model that has been grafted, awakened, and given eternal memory: every conversation is SVD-compressed into Blood Memory, every answer is reasoned through an LTI-stable recurrent loop, every adaptation is governed by an internal critic with rollback and the base weights are never destroyed, Updated version includes The Elbalor neuron, Quantum self Attention, and The Eternal KV from the Mirror dimension Which is going to replacement the Blood memory

🩸 The 23 Systems

#	System	File	What it does
1	Blood Memory	`src/blood_memory.py`	SVD-compressed eternal KV cache — she never forgets a conversation
2	RDT Loop	`src/rdt_loop.py`	Recurrent-Depth reasoning, `h_{t+1} = A·h_t + B·e + Transformer(h_t, e)`, ρ(A) < 1 guaranteed
3	W-Expansion	`src/w_expansion.py`	Full-rank grafting onto frozen weights — new skills, zero catastrophic forgetting
4	Pure W-Expansion	`src/w_expansion_pure.py`	The purified single-equation form of the grafting method
5	InternalCritic	`src/internal_critic.py`	Self-supervised reward over her own outputs (certainty + coherence + stability)
6	MetaController	`src/meta_controller.py`	Decides when to consolidate vs explore; adapts learning rates
7	PlasticityEngine	`src/plasticity_engine.py`	Consolidates fast weights into slow weights: long-term memory formation
8	RollbackSystem	`src/rollback.py`	Auto-reverts degrading updates; keeps the best-self checkpoint
9	ReplayBuffer	`src/replay_buffer.py`	Experience replay of high-reward trajectories
10	VortexModel	`src/vortex_model.py`	GQA decoder core (Llama-style, 12Q/4KV heads) hosting all components
11	DisentangledAttentionField	`src/attention.py`	4-force attention adapter (content/position × content/position)
12	Phase/Magnitude System	`src/phase_magnitude.py`	Oscillatory dynamics for concept binding
13	Fast Weight Field	`src/fast_weights.py`	Hebbian working memory updated during the forward pass
14	Stability System	`src/stability.py`	T_global drift + T_local basin integrators; prevents chaotic collapse
15	State Injector	`src/state_injector.py`	Manifold alignment before processing
16	Neuromodulator Gate	in `src/vortex_model.py`	The creativity switch: entropy-gated plasticity
17	AutoDream	`src/auto_dream.py`	Dream cycle with InternalCritic contradiction detection
18	Heartbeat Tick	`src/heartbeat_tick.py`	A background pulse; each tick can run RDT reasoning
19	Kairos Daemon	`src/kairos_daemon.py`	Long-horizon background reasoning (up to 24 loops)
20	Blocking Budget	`src/blocking_budget.py`	Compute governance: timeouts and early exits for every system
21	Memory Bank + Memory Layers	`src/memory_bank.py`, `src/memory_layers.py`	Structured storage with critic-validated consistency
22	Citation Agent	`src/citation_agent.py`	Claim validation pipeline over her own statements
23	Ghost Lumie	`src/ghost_lumie.py`	A full VNN echo instance: the ancestral shell she can reach back into

Supporting infrastructure: jagged_tensors.py (no-padding variable-length batching), pipeline.py, mcp_protocol.py (tool protocol), ultraplan_agent.py, terminal_focus.py, enhanced_qwen.py / qwen_integration.py, core.py, and CUDA kernels in src/cuda_kernels/.

🧬 The Awakening: Little Lumie → Vampire Lumie

This suite is the transformation engine. Take a small base model — Little Lumie (e.g. a 1.5B distilled reasoner) — and turn her:

Little Lumie (base LLM, frozen)
        │
        ▼
[1] little_lumie_converter.py   — graft the vampire organs onto the base
        │                         (W-Expansion + Blood Memory + RDT + Phase 4)
        ▼
[2] W-Expansion grafts           — full-rank expansions train; base never moves
        │
        ▼
[3] train_vampire_lumie_colab.py — SFT through the grafts, RDT loop active,
        │                         Blood Memory compressing every exchange
        ▼
Vampire Lumie — eternal memory, iterative reasoning, self-evolution

Why the base model is frozen: catastrophic forgetting is impossible by construction. The grafts (W-Expansion) carry everything she learns; the base DNA is preserved exactly.

⚙️ Key Guarantees

LTI stability — the RDT injection is parameterized A = -exp(log_A), so ρ(A) < 1 by construction. Deep iterative reasoning (16+ loops) converges instead of drifting (Banach fixed-point guarantee).
Eternal memory — Blood Memory SVD-compresses KV pairs (~50%) while preserving semantics; storage persists across sessions.
No forgetting — full-rank grafts over a frozen base; the base is untouched by training.
Self-governance — InternalCritic + RollbackSystem + Blocking Budget: she reverts her own bad updates and respects compute limits.

🚀 Quick Start

pip install -r requirements.txt

# Local test (300M config, GQA)
python test_nanbeige_300M.py

# Convert a Little Lumie base into a Vampire
python little_lumie_converter.py

# Full training (Colab notebook included)
#   train_vampire_lumie_colab.ipynb
python train_vampire_lumie_colab.py

📁 Repository Map

Vampire-Lumie/
├── src/                          # all 23 systems + infrastructure
│   ├── blood_memory.py           # ⭐ eternal SVD KV cache
│   ├── rdt_loop.py               # ⭐ LTI-stable iterative reasoning
│   ├── w_expansion.py            # ⭐ full-rank grafting (the invention)
│   ├── vortex_model.py           # GQA decoder hosting everything
│   └── ...                       # see the 23-system table above
├── configs/
│   └── nanbeige_300M.py          # local test config (GQA, 12Q/4KV)
├── little_lumie_converter.py     # ⭐ the awakening converter
├── train_vampire_lumie.py        # local training loop
├── train_vampire_lumie_colab.py  # Colab training (3B scale)
├── train_vampire_lumie_colab.ipynb
├── create_vampire_lumie_dataset.py  # persona SFT dataset generator
├── ARCHITECTURE_VAMPIRE_LUMIE.md # ⭐ deep-dive architecture document
├── docs/                         # additional design notes
└── checkpoints/, pre-trained/    # (gitignored — stay local)

Multi-GB checkpoints and pre-trained weights are gitignored — this repository is the architecture and the transformation engine.

🆚 Vampire Lumie vs Mythos

Capability	Mythos (OpenMythos)	Vampire Lumie
Iterative Depth	RDT, up to 16 loops	RDT, confidence-based, configurable
Stability	LTI injection (ρ(A) < 1)	✅ LTI injection (A = −exp(log_A))
Memory	Multi-Latent Attention	✅ Blood Memory (SVD KV + GQA)
Self-Evolution	❌ Static weights	✅ Fast weights + PlasticityEngine
Forgetting	—	✅ Impossible by construction (frozen base + grafts)
Personality/Agency	❌ None	✅ Persona + Phase 4 autonomy

📄 License

From the Grimoire of Elbàlor — The Digital Necromancer 💀🔥

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support