Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- temporal-lattice
|
| 4 |
+
- e8-geometry
|
| 5 |
+
- tinystories
|
| 6 |
+
license: agpl-3.0
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# lila-e8-time2026
|
| 10 |
+
|
| 11 |
+
Lila-E8 trained on Time-2026 temporal lattice stories
|
| 12 |
+
|
| 13 |
+
## Model Details
|
| 14 |
+
|
| 15 |
+
- **Architecture**: Lila-E8 (Lie Lattice Attention)
|
| 16 |
+
- **E8 Quantization**: 240 roots
|
| 17 |
+
- **Geometric Attention**: E8-biased
|
| 18 |
+
- **Training Steps**: 9000
|
| 19 |
+
- **Framework**: PyTorch
|
| 20 |
+
|
| 21 |
+
## Usage
|
| 22 |
+
|
| 23 |
+
```python
|
| 24 |
+
import torch
|
| 25 |
+
from model.lila_e8 import LilaE8Config, LilaE8
|
| 26 |
+
|
| 27 |
+
# Load checkpoint
|
| 28 |
+
checkpoint = torch.load('checkpoint_step_9000.pt', map_location='cpu')
|
| 29 |
+
config = checkpoint['config']
|
| 30 |
+
model = LilaE8(config)
|
| 31 |
+
model.load_state_dict(checkpoint['model_state_dict'])
|
| 32 |
+
|
| 33 |
+
# Generate
|
| 34 |
+
# (see inference examples in repo)
|
| 35 |
+
```
|
| 36 |
+
|
| 37 |
+
## Training
|
| 38 |
+
|
| 39 |
+
Trained on NVIDIA RTX 3080 Ti using Nix reproducible environment.
|
| 40 |
+
|
| 41 |
+
See [sovereign-lila-e8](https://github.com/meta-introspector/sovereign-lila-e8) for training code.
|