Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
pipeline_tag: text-generation
|
| 6 |
+
tags:
|
| 7 |
+
- transformer
|
| 8 |
+
- gpt2
|
| 9 |
+
- tiny
|
| 10 |
+
- atom
|
| 11 |
+
- experimental
|
| 12 |
+
- humor
|
| 13 |
+
- whirlwindai
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
<div align="center">
|
| 17 |
+
|
| 18 |
+
<img src="https://readme-typing-svg.demolab.com?font=Space+Grotesk&weight=700&size=27&duration=1900&pause=900&color=22C55E¢er=true&vCenter=true&width=850&lines=AtomZephyr;27+Parameters.;Almost+an+LLM.;Mostly+a+Science+Experiment.;Powered+by+Pure+Curiosity." />
|
| 19 |
+
|
| 20 |
+
<br>
|
| 21 |
+
|
| 22 |
+
<img src="https://img.shields.io/badge/Parameters-27-22C55E?style=for-the-badge">
|
| 23 |
+
<img src="https://img.shields.io/badge/Architecture-GPT--2-10B981?style=for-the-badge">
|
| 24 |
+
<img src="https://img.shields.io/badge/Status-Experimental-14B8A6?style=for-the-badge">
|
| 25 |
+
<img src="https://img.shields.io/badge/Braincells-27-06B6D4?style=for-the-badge">
|
| 26 |
+
|
| 27 |
+
<br><br>
|
| 28 |
+
|
| 29 |
+
<img src="https://capsule-render.vercel.app/api?type=soft&height=190&text=AtomZephyr&fontColor=ffffff&fontSize=48&animation=blinking&color=0:22C55E,100:06B6D4"/>
|
| 30 |
+
|
| 31 |
+
</div>
|
| 32 |
+
|
| 33 |
+
---
|
| 34 |
+
|
| 35 |
+
# The Idea
|
| 36 |
+
|
| 37 |
+
<div align="center">
|
| 38 |
+
|
| 39 |
+
<table width="92%">
|
| 40 |
+
|
| 41 |
+
<tr>
|
| 42 |
+
|
| 43 |
+
<td align="center">
|
| 44 |
+
|
| 45 |
+
## What if a transformer became... microscopic?
|
| 46 |
+
|
| 47 |
+
AtomZephyr explores one of the smallest practical transformer architectures ever built.
|
| 48 |
+
|
| 49 |
+
Not because anyone asked for it.
|
| 50 |
+
|
| 51 |
+
Because someone eventually had to answer the question:
|
| 52 |
+
|
| 53 |
+
**"How absurdly small can an AI become before it forgets how to AI?"**
|
| 54 |
+
|
| 55 |
+
Turns out...
|
| 56 |
+
|
| 57 |
+
**27 parameters is still technically enough.**
|
| 58 |
+
|
| 59 |
+
</td>
|
| 60 |
+
|
| 61 |
+
</tr>
|
| 62 |
+
|
| 63 |
+
</table>
|
| 64 |
+
|
| 65 |
+
</div>
|
| 66 |
+
|
| 67 |
+
---
|
| 68 |
+
|
| 69 |
+
# Why?
|
| 70 |
+
|
| 71 |
+
Most AI models compete by getting bigger.
|
| 72 |
+
|
| 73 |
+
AtomZephyr competes by removing parameters until people start questioning whether it's still a neural network.
|
| 74 |
+
|
| 75 |
+
Every parameter had to earn its place.
|
| 76 |
+
|
| 77 |
+
Most didn't.
|
| 78 |
+
|
| 79 |
+
---
|
| 80 |
+
|
| 81 |
+
# Specifications
|
| 82 |
+
|
| 83 |
+
| Property | Value |
|
| 84 |
+
|-----------|-------|
|
| 85 |
+
| Parameters | **27** |
|
| 86 |
+
| Architecture | GPT-2 |
|
| 87 |
+
| Layers | 1 |
|
| 88 |
+
| Attention Heads | 1 |
|
| 89 |
+
| Embedding Size | 1 |
|
| 90 |
+
| FFN Size | 1 |
|
| 91 |
+
| Context Length | 4 |
|
| 92 |
+
| Vocabulary | 5 Tokens |
|
| 93 |
+
| Model Size | <5 KB |
|
| 94 |
+
| Training Time | ~6 Seconds (CPU) |
|
| 95 |
+
|
| 96 |
+
---
|
| 97 |
+
|
| 98 |
+
# Performance
|
| 99 |
+
|
| 100 |
+
| Test | Result |
|
| 101 |
+
|------|--------|
|
| 102 |
+
| Understand English | β |
|
| 103 |
+
| Write Code | β |
|
| 104 |
+
| Solve Math | β |
|
| 105 |
+
| Generate "abba" | β
|
|
| 106 |
+
| Break Expectations | β
|
|
| 107 |
+
|
| 108 |
+
---
|
| 109 |
+
|
| 110 |
+
# Quick Start
|
| 111 |
+
|
| 112 |
+
```python
|
| 113 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 114 |
+
|
| 115 |
+
tokenizer = AutoTokenizer.from_pretrained("WhirlwindAI/AtomZephyr")
|
| 116 |
+
model = AutoModelForCausalLM.from_pretrained("WhirlwindAI/AtomZephyr")
|
| 117 |
+
|
| 118 |
+
prompt = "a"
|
| 119 |
+
|
| 120 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
| 121 |
+
|
| 122 |
+
outputs = model.generate(
|
| 123 |
+
**inputs,
|
| 124 |
+
do_sample=True,
|
| 125 |
+
temperature=1.7,
|
| 126 |
+
max_length=4
|
| 127 |
+
)
|
| 128 |
+
|
| 129 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
| 130 |
+
```
|
| 131 |
+
|
| 132 |
+
Possible output
|
| 133 |
+
|
| 134 |
+
```
|
| 135 |
+
abaa
|
| 136 |
+
```
|
| 137 |
+
|
| 138 |
+
Groundbreaking.
|
| 139 |
+
|
| 140 |
+
---
|
| 141 |
+
|
| 142 |
+
# Example Conversation
|
| 143 |
+
|
| 144 |
+
**User**
|
| 145 |
+
|
| 146 |
+
> Tell me a joke.
|
| 147 |
+
|
| 148 |
+
**AtomZephyr**
|
| 149 |
+
|
| 150 |
+
```
|
| 151 |
+
abba
|
| 152 |
+
```
|
| 153 |
+
|
| 154 |
+
Technically...
|
| 155 |
+
|
| 156 |
+
that's an answer.
|
| 157 |
+
|
| 158 |
+
---
|
| 159 |
+
|
| 160 |
+
# Scientific Achievement
|
| 161 |
+
|
| 162 |
+
Removing parameters is easy.
|
| 163 |
+
|
| 164 |
+
Keeping a transformer alive afterwards...
|
| 165 |
+
|
| 166 |
+
isn't.
|
| 167 |
+
|
| 168 |
+
AtomZephyr exists purely to explore the absolute lower limits of transformer architectures while remaining a real, trainable language model.
|
| 169 |
+
|
| 170 |
+
Whether it's useful is a completely different discussion.
|
| 171 |
+
|
| 172 |
+
---
|
| 173 |
+
|
| 174 |
+
# Awards
|
| 175 |
+
|
| 176 |
+
π₯ Smallest Model That Still Has Self-Respect
|
| 177 |
+
|
| 178 |
+
π Best Binary Poetry Generator
|
| 179 |
+
|
| 180 |
+
π₯ Most Efficient Waste Of Six Seconds
|
| 181 |
+
|
| 182 |
+
ποΈ Official Representative Of Tiny AI
|
| 183 |
+
|
| 184 |
+
---
|
| 185 |
+
|
| 186 |
+
# Limitations
|
| 187 |
+
|
| 188 |
+
AtomZephyr should **not** be used for:
|
| 189 |
+
|
| 190 |
+
- Programming
|
| 191 |
+
- Translation
|
| 192 |
+
- Question Answering
|
| 193 |
+
- Homework
|
| 194 |
+
- Anything important
|
| 195 |
+
|
| 196 |
+
It performs significantly better when asked to do absolutely nothing useful.
|
| 197 |
+
|
| 198 |
+
---
|
| 199 |
+
|
| 200 |
+
# Fun Facts
|
| 201 |
+
|
| 202 |
+
- Fits inside most PNG images.
|
| 203 |
+
- Smaller than many neural network tutorials.
|
| 204 |
+
- Downloads faster than this README loads.
|
| 205 |
+
- Has fewer parameters than some calculator manuals have pages.
|
| 206 |
+
|
| 207 |
+
---
|
| 208 |
+
|
| 209 |
+
# License
|
| 210 |
+
|
| 211 |
+
MIT
|
| 212 |
+
|
| 213 |
+
Take it apart.
|
| 214 |
+
|
| 215 |
+
Make it smaller.
|
| 216 |
+
|
| 217 |
+
Break another record.
|
| 218 |
+
|
| 219 |
+
---
|
| 220 |
+
|
| 221 |
+
<div align="center">
|
| 222 |
+
|
| 223 |
+
### Built by WhirlwindAI
|
| 224 |
+
|
| 225 |
+
*"Sometimes progress isn't measured in billions... it's measured in what you can remove."*
|
| 226 |
+
|
| 227 |
+
<br>
|
| 228 |
+
|
| 229 |
+
<img src="https://capsule-render.vercel.app/api?type=waving&height=120§ion=footer&color=0:22C55E,100:06B6D4"/>
|
| 230 |
+
|
| 231 |
+
</div>
|
| 232 |
+
|