docs: update README
Browse files
README.md
CHANGED
@@ -1,3 +1,79 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- ru
|
5 |
+
tags:
|
6 |
+
- distill
|
7 |
+
- fill-mask
|
8 |
+
- embeddings
|
9 |
+
- masked-lm
|
10 |
+
- tiny
|
11 |
+
- feature-extraction
|
12 |
+
- sentence-similarity
|
13 |
+
datasets:
|
14 |
+
- GEM/wiki_lingua
|
15 |
+
- xnli
|
16 |
+
- RussianNLP/wikiomnia
|
17 |
+
- mlsum
|
18 |
+
- IlyaGusev/gazeta
|
19 |
+
widget:
|
20 |
+
- text: Москва - <mask> России.
|
21 |
+
- text: Если б море было пивом, я бы <mask>
|
22 |
+
- text: Столица России - <mask>.
|
23 |
---
|
24 |
+
# ruRoberta-distilled
|
25 |
+
|
26 |
+
Model was distilled from [ai-forever/ruRoberta-large](https://huggingface.co/ai-forever/ruRoberta-large) with ❤️ by me for 120 hours using 4 Nvidia V100.
|
27 |
+
|
28 |
+
## Usage
|
29 |
+
|
30 |
+
```python
|
31 |
+
from transformers import pipeline
|
32 |
+
|
33 |
+
|
34 |
+
pipe = pipeline('feature-extraction', model='d0rj/ruRoberta-distilled')
|
35 |
+
tokens_embeddings = pipe('Привет, мир!')
|
36 |
+
```
|
37 |
+
|
38 |
+
```python
|
39 |
+
import torch
|
40 |
+
from transformers import AutoTokenizer, AutoModel
|
41 |
+
|
42 |
+
|
43 |
+
tokenizer = AutoTokenizer.from_pretrained('d0rj/ruRoberta-distilled')
|
44 |
+
model = AutoModel.from_pretrained('d0rj/ruRoberta-distilled')
|
45 |
+
|
46 |
+
|
47 |
+
def embed_bert_cls(text: str) -> torch.Tensor:
|
48 |
+
t = tokenizer(text, padding=True, truncation=True, return_tensors='pt').to(model.device)
|
49 |
+
with torch.no_grad():
|
50 |
+
model_output = model(**t)
|
51 |
+
embeddings = model_output.last_hidden_state[:, 0, :]
|
52 |
+
embeddings = torch.nn.functional.normalize(embeddings)
|
53 |
+
return embeddings[0].cpu()
|
54 |
+
|
55 |
+
|
56 |
+
embedding = embed_bert_cls('Привет, мир!')
|
57 |
+
```
|
58 |
+
|
59 |
+
## Logs
|
60 |
+
|
61 |
+
See all logs at [WandB](https://wandb.ai/d0rj/distill-ruroberta/runs/lehtr3bk/workspace).
|
62 |
+
|
63 |
+
## Configuration
|
64 |
+
|
65 |
+
- Activation GELU -> GELUFast
|
66 |
+
- Attention heads 16 -> 8
|
67 |
+
- Hidden layers 24 -> 6
|
68 |
+
- Weights size 1.42 GB -> 464 MB
|
69 |
+
|
70 |
+
## Data
|
71 |
+
|
72 |
+
Overall: 9.4 GB of raw texts, 5.1 GB of binarized texts.
|
73 |
+
|
74 |
+
Used data:
|
75 |
+
- [GEM/wiki_lingua](https://huggingface.co/datasets/GEM/wiki_lingua)
|
76 |
+
- [xnli](https://huggingface.co/datasets/xnli)
|
77 |
+
- [RussianNLP/wikiomnia](https://huggingface.co/datasets/RussianNLP/wikiomnia)
|
78 |
+
- [mlsum](https://huggingface.co/datasets/mlsum)
|
79 |
+
- [IlyaGusev/gazeta](https://huggingface.co/datasets/IlyaGusev/gazeta)
|