Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- gpt
|
| 4 |
+
- from-scratch
|
| 5 |
+
- code-generation
|
| 6 |
+
- rope
|
| 7 |
+
- gqa
|
| 8 |
+
- swiglu
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# gpt-python-modern-114m
|
| 12 |
+
|
| 13 |
+
A ~114M parameter LLaMA/Mistral-style decoder-only transformer (RMSNorm, RoPE, SwiGLU,
|
| 14 |
+
Grouped Query Attention), trained from scratch on
|
| 15 |
+
[devoppro/Python](https://huggingface.co/datasets/devoppro/Python).
|
| 16 |
+
|
| 17 |
+
- Layers: 12, Query heads: 12, KV heads: 4
|
| 18 |
+
- Embedding dim: 768
|
| 19 |
+
- Context length: 512
|
| 20 |
+
- Trained to step 0, val loss 10.3440
|
| 21 |
+
- Tokenizer: GPT-2 BPE (tiktoken 'gpt2')
|