✦ Veytra ✦
From words to vectors, from vectors to meaning.
A lightweight, elegant sentence embedding model, built from scratch.
Trained with a Transformer architecture, Veytra maps sentences into 64-dimensional vectors and measures the semantic closeness between two sentences via cosine similarity. Small yet ambitious — designed for those who believe in the power of simplicity.
| ⚙️ Architecture | |
|---|---|
| Total Parameters | ~3.3M (3,316,544) |
| Tokenizer | GPT-2 (50,257 vocab) |
| Model | Transformer Encoder (2 layers, 4 heads) |
| Embedding Dimension | 64 |
| Max Length | 64 tokens |
| Pooling | Mean Pooling + L2 Normalization |
⚡ Usage
python3 train.py
Veytra — encoding meaning.