Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- pytorch
|
| 5 |
+
- causal-lm
|
| 6 |
+
- mixture-of-experts
|
| 7 |
+
- from-scratch
|
| 8 |
+
- intelligence-os
|
| 9 |
+
pipeline_tag: text-generation
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Intelligence_OS (1.5 Billion Parameter Sparse Architecture)
|
| 13 |
+
|
| 14 |
+
Intelligence_OS is a decoder-only sparse mixture-of-experts language model designed from scratch for maximum quality per joule and compute-efficient training.
|
| 15 |
+
|
| 16 |
+
## Model Details
|
| 17 |
+
- **Architecture**: Decoder-only Sparse MoE
|
| 18 |
+
- **Parameters**: 1,518,659,584 (theoretical 1.519B class)
|
| 19 |
+
- **Layers**: 26
|
| 20 |
+
- **Hidden Dimension**: 2,048
|
| 21 |
+
- **Attention**: Grouped-Query Attention (16 Q heads / 4 KV heads)
|
| 22 |
+
- **Position Encoding**: Rotary Position Embedding (RoPE)
|
| 23 |
+
- **Feed-Forward**: 4 SwiGLU Experts per layer with Top-2 and Adaptive Routing
|
| 24 |
+
- **Embeddings**: Tied input/output token embeddings
|
| 25 |
+
|
| 26 |
+
## Training Innovations
|
| 27 |
+
- **Low-Memory Optimization (LOMO)**: Layer-wise in-place updates keeping peak VRAM under 1.5 GB.
|
| 28 |
+
- **Zero-GPU Non-Neural AI Engine**: Holographic Associative Memory capable of 85k tokens/s.
|