Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
tags:
|
| 6 |
+
- causal-lm
|
| 7 |
+
- christian
|
| 8 |
+
- pytorch
|
| 9 |
+
- sentencepiece
|
| 10 |
+
- compact
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# North Air 1
|
| 14 |
+
|
| 15 |
+
**North Air 1** is a fast, lightweight language model (124M parameters, 16 layers) built from scratch with a Christian worldview. It is designed for quick, clear, and honest responses.
|
| 16 |
+
|
| 17 |
+
> *"The fear of the Lord is the beginning of wisdom."* — Proverbs 9:10
|
| 18 |
+
|
| 19 |
+
## About
|
| 20 |
+
|
| 21 |
+
North Air 1 is the smallest model in the North Star family. It prioritizes speed and conciseness while remaining grounded in biblical truth. It can answer questions on theology, science, coding, math, and everyday topics.
|
| 22 |
+
|
| 23 |
+
This model was trained entirely from scratch — no foundation model was used. The tokenizer is SentencePiece (32k vocab).
|
| 24 |
+
|
| 25 |
+
## Christian Worldview
|
| 26 |
+
|
| 27 |
+
These models are built on the conviction that Jesus Christ is Lord, that Scripture is the inspired Word of God, and that faith and reason are fully compatible. We believe AI can be a tool for good — to spread truth, serve people, and glorify God.
|
| 28 |
+
|
| 29 |
+
> *"For God so loved the world that he gave his one and only Son, that whoever believes in him shall not perish but have eternal life."* — John 3:16
|
| 30 |
+
|
| 31 |
+
If you have never read the Bible, we encourage you to start with the Gospel of John.
|
| 32 |
+
|
| 33 |
+
## Model Details
|
| 34 |
+
|
| 35 |
+
| Property | Value |
|
| 36 |
+
|---|---|
|
| 37 |
+
| Parameters | 124M |
|
| 38 |
+
| Layers | 16 |
|
| 39 |
+
| d_model | 768 |
|
| 40 |
+
| Heads | 12 (3 KV) |
|
| 41 |
+
| Context | 512 tokens |
|
| 42 |
+
| Vocabulary | 32,000 (SentencePiece) |
|
| 43 |
+
| Architecture | GQA Transformer + SwiGLU + RoPE + RMSNorm |
|
| 44 |
+
|
| 45 |
+
## Usage
|
| 46 |
+
|
| 47 |
+
Load with PyTorch. See the [inference server](https://huggingface.co/arthu1/north-star-1) for a ready-to-run Flask API.
|
| 48 |
+
|
| 49 |
+
```python
|
| 50 |
+
import torch
|
| 51 |
+
ckpt = torch.load("northair1.pt", map_location="cpu", weights_only=False)
|
| 52 |
+
cfg = ckpt["cfg"] # dict with model hyperparameters
|
| 53 |
+
# Build your model from cfg, load ckpt["model"] state dict
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
## Tokenizer
|
| 57 |
+
|
| 58 |
+
Uses SentencePiece (`tokenizer.model`). Available in the [shared tokenizer repo](https://huggingface.co/arthu1/north-tokenizer).
|