Update README.md
Browse files
README.md
CHANGED
|
@@ -3,4 +3,19 @@ license: mit
|
|
| 3 |
datasets:
|
| 4 |
- allenai/c4
|
| 5 |
pipeline_tag: text-generation
|
| 6 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
datasets:
|
| 4 |
- allenai/c4
|
| 5 |
pipeline_tag: text-generation
|
| 6 |
+
---
|
| 7 |
+
Wigip-1: A 473M Parameter Language Model
|
| 8 |
+
This repository contains the code and documentation for Wigip-1, a ~500M parameter GPT-style language model built from scratch in JAX/Flax.
|
| 9 |
+
|
| 10 |
+
Project Overview
|
| 11 |
+
This project was an end-to-end journey into building and training a large language model on public resources. It involved:
|
| 12 |
+
|
| 13 |
+
Architecture: A 24-layer, 1280-embedding dimension Transformer.
|
| 14 |
+
Training: Trained on the C4 dataset for over 500,000 steps (~8 hours on a TPU v3-8).
|
| 15 |
+
Frameworks: Built with JAX, Flax, and Optax.
|
| 16 |
+
Deployment: A live demo was created using Gradio.
|
| 17 |
+
The trained model weights are hosted separately on the Hugging Face Hub, as they are too large for a standard Git repository:
|
| 18 |
+
https://huggingface.co/Nottybro/wigip-1
|
| 19 |
+
|
| 20 |
+
My Journey
|
| 21 |
+
This project was a deep dive into the real-world challenges of MLOps, including debugging file corruption, solving JAX compiler errors (XlaRuntimeError), and managing long-running jobs in a cloud environment. It was built with the help of an AI assistant for debugging and guidance.
|