Metadata-Version: 2.4
Name: rl2t
Version: 0.7.0
Summary: RL2T: Recursive Liquid-Lattice Text Architecture — Self-Expanding Non-Transformer Deep RL Neural Network with YaRN-RoPE, Thalamo-Cortical Global Workspace, and Universal Model Transmutation (.gguf, .safetensors, .pt, .npz)
Author: PlimbAi
Author-email: PlimbAi <plimb.ai.contact@gmail.com>
Maintainer-email: PlimbAi <plimb.ai.contact@gmail.com>
License-Expression: MIT
Keywords: rl2t,liquid-neural-networks,reinforcement-learning,neurogenesis,rope,gguf,safetensors,transformer-transmutation,plimbai
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: Programming Language :: Python :: 3
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: torch>=2.0.0
Requires-Dist: numpy>=1.22.0
Dynamic: author
Dynamic: license-file
Dynamic: requires-python

# RL2T — Recursive Liquid-Lattice Text Architecture (v0.7.0)

**Author:** PlimbAi (`plimb.ai.contact@gmail.com`)  
**Package:** `rl2t`

**RL2T** is a self-expanding, non-Transformer Deep Reinforcement Learning neural architecture equipped with:
- **Multi-Scale YaRN-RoPE (`ExpandableRotaryEmbedding`)** with Hippocampal Phase Precession & automatic width neurogenesis
- **Thalamo-Cortical Global Workspace Bus (`ThalamoCorticalGlobalWorkspace`)** integrating all cortical layers simultaneously
- **Dual-Timescale Liquid State-Space Cells (`LiquidTimeConstantCell`)** with RK2 second-order refinement, Targeted Neuron Mitosis, and Synaptic Consolidation
- **Rotary Dual-Bank Holographic Associative Memory (`HolographicAssociativeMemory`)**
- **4-Hormone Computational Endocrine System (`NeuromodulatorSystem`)**: Dopamine (`DA`), Acetylcholine (`ACh`), Norepinephrine (`NE`), and Serotonin (`5-HT`)
- **Universal Multi-Format Transmuter (`RL2TTransmuter`)**: Import and surgically transmute any Transformer model (`.gguf`, `.safetensors`, `.pt`, `.pth`, `.bin`, `.ckpt`, `.npz`, `.npy`, HuggingFace folder, or PyTorch `nn.Module`) into RL2T!
- **Autonomous Out-of-Training Growth**: Train normally (`fit_normal`), then execute anywhere and watch the model continue growing its weights and architecture automatically (`stop_growth=False`), or lock its size at any time with `stop_growth=True`.

---

## Installation

```bash
pip install rl2t
```

---

## Quick Start

```python
import torch
from rl2t import RL2TModel, RL2TTransmuter

# 1. Start from ANY scale ('femto', 'pico', 'nano', 'micro', 'small', 'medium', 'large', or exact parameter budget)
model = RL2TModel.from_any_scale("nano", vocab_size=256, stop_growth=False)

# 2. Or import & transmute ANY existing model (.gguf, .safetensors, .pt, .npz, or PyTorch nn.Module)
# model, report = RL2TTransmuter.from_any_file("model.gguf", stop_growth=False)

# 3. Normal supervised training (also calibrates internal RL Critic & World-Model)
batch = torch.randint(0, 256, (4, 32))
model.fit_normal(batch, epochs=5, lr=3e-3, allow_growth_during_training=False)

# 4. Execute outside training -> continues growing autonomously unless stop_growth=True!
out = model(batch[:2])
print("Current parameters:", model.count_parameters())

# 5. Lock growth whenever desired
model.set_growth_lock(stop_growth=True)
```

---

## Contact & License

- **Author:** PlimbAi
- **Email:** plimb.ai.contact@gmail.com
- **License:** MIT
