Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
library_name: julia
|
| 5 |
+
pipeline_tag: text-generation
|
| 6 |
+
tags:
|
| 7 |
+
- character-level
|
| 8 |
+
- philosophy
|
| 9 |
+
- mathematics
|
| 10 |
+
- julia
|
| 11 |
+
- scalar-autograd
|
| 12 |
+
- pure-julia
|
| 13 |
+
datasets:
|
| 14 |
+
- LisaMegaWatts/microjulia-data
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# MicroJulia
|
| 18 |
+
|
| 19 |
+
A minimal character-level GPT built entirely in pure Julia with scalar autograd. No external ML dependencies.
|
| 20 |
+
|
| 21 |
+
## Architecture
|
| 22 |
+
- 1 transformer layer, 4 attention heads
|
| 23 |
+
- n_embd=16, block_size=64
|
| 24 |
+
- RMSNorm, ReLU, KV cache for causal masking
|
| 25 |
+
- Adam optimizer with linear LR decay
|
| 26 |
+
- ~5K parameters
|
| 27 |
+
|
| 28 |
+
## Vocabulary
|
| 29 |
+
27 characters (a-z + space) + BOS = 28 vocab
|
| 30 |
+
|
| 31 |
+
## Training
|
| 32 |
+
- **Dataset:** Aristotle's Rhetoric + Euclid's Elements (8,487 chunks)
|
| 33 |
+
- **Current checkpoint:** step 150, val_loss=2.4315
|
| 34 |
+
|
| 35 |
+
## Links
|
| 36 |
+
- [Live inference (HF Space)](https://huggingface.co/spaces/LisaMegaWatts/MicroJulia)
|
| 37 |
+
- [Training data](https://huggingface.co/datasets/LisaMegaWatts/microjulia-data)
|
| 38 |
+
- [Source code](https://github.com/DavinciDreams/micro-julia)
|