Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- text-embeddings
|
| 4 |
+
- jepa
|
| 5 |
+
- self-supervised
|
| 6 |
+
- vicreg
|
| 7 |
+
library_name: pytorch
|
| 8 |
+
license: mit
|
| 9 |
+
source_code: https://github.com/vasil646srb/jepa_text
|
| 10 |
+
datasets:
|
| 11 |
+
- HuggingFaceFW/fineweb-edu
|
| 12 |
+
base_model:
|
| 13 |
+
- BAAI/bge-small-en-v1.5
|
| 14 |
+
language:
|
| 15 |
+
- en
|
| 16 |
+
metrics:
|
| 17 |
+
- cosine-similarity
|
| 18 |
+
pipeline_tag: feature-extraction
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# Text-JEPA Streaming Embeddings
|
| 22 |
+
|
| 23 |
+
Семантические эмбеддинги, обученные с помощью **Text-JEPA** (Joint Embedding Predictive Architecture) на датасете FineWeb-Edu через Streaming Pipeline.
|
| 24 |
+
|
| 25 |
+
## 🚀 Особенности
|
| 26 |
+
|
| 27 |
+
- **Базовая модель**: BGE-Small-EN-v1.5 (ONNX)
|
| 28 |
+
- **Архитектура**: 2-layer Transformer Encoder + VICReg Loss
|
| 29 |
+
- **Output dimension**: 128
|
| 30 |
+
- **Поддержка**: NVIDIA GPU (CUDA) и CPU (AVX2/FMA) с автоопределением
|
| 31 |
+
- **Обучение**: 50K+ шагов на 500K примерах (Streaming)
|
| 32 |
+
|
| 33 |
+
## 📦 Быстрый старт
|
| 34 |
+
|
| 35 |
+
```python
|
| 36 |
+
from huggingface_hub import snapshot_download
|
| 37 |
+
|
| 38 |
+
# Скачать все веса и код
|
| 39 |
+
snapshot_download("vasil646/jepa_text", local_dir="./jepa_text")
|