Commit
•
8084f28
1
Parent(s):
0eb6f15
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
datasets:
|
| 4 |
+
- squad_v2
|
| 5 |
+
license: mit
|
| 6 |
+
thumbnail: https://thumb.tildacdn.com/tild3433-3637-4830-a533-353833613061/-/resize/720x/-/format/webp/germanquad.jpg
|
| 7 |
+
tags:
|
| 8 |
+
- exbert
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
## Overview
|
| 12 |
+
**Language model:** deepset/tinybert-6L-768D-squad2
|
| 13 |
+
**Language:** English
|
| 14 |
+
**Training data:** SQuAD 2.0 training set x 20 augmented + SQuAD 2.0 training set
|
| 15 |
+
**Eval data:** SQuAD 2.0 dev set
|
| 16 |
+
**Infrastructure**: 1x V100 GPU
|
| 17 |
+
**Published**: Dec 8th, 2021
|
| 18 |
+
|
| 19 |
+
## Details
|
| 20 |
+
- haystack's intermediate layer and prediction layer distillation features were used for training (based on [TinyBERT](https://arxiv.org/pdf/1909.10351.pdf)). deepset/bert-base-uncased-squad2 was used as the teacher model.
|
| 21 |
+
|
| 22 |
+
## Hyperparameters
|
| 23 |
+
### Intermediate layer distillation
|
| 24 |
+
```
|
| 25 |
+
batch_size = 26
|
| 26 |
+
n_epochs = 5
|
| 27 |
+
max_seq_len = 384
|
| 28 |
+
learning_rate = 5e-5
|
| 29 |
+
lr_schedule = LinearWarmup
|
| 30 |
+
embeds_dropout_prob = 0.1
|
| 31 |
+
temperature = 1
|
| 32 |
+
distillation_loss_weight = 0.75
|
| 33 |
+
```
|
| 34 |
+
### Prediction layer distillation
|
| 35 |
+
```
|
| 36 |
+
batch_size = 26
|
| 37 |
+
n_epochs = 5
|
| 38 |
+
max_seq_len = 384
|
| 39 |
+
learning_rate = 3e-5
|
| 40 |
+
lr_schedule = LinearWarmup
|
| 41 |
+
embeds_dropout_prob = 0.1
|
| 42 |
+
temperature = 1
|
| 43 |
+
distillation_loss_weight = 0.75
|
| 44 |
+
```
|
| 45 |
+
## Performance
|
| 46 |
+
```
|
| 47 |
+
"exact": 71.87736882001179
|
| 48 |
+
"f1": 76.36111895973675
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
## Authors
|
| 52 |
+
- Timo Möller: `timo.moeller [at] deepset.ai`
|
| 53 |
+
- Julian Risch: `julian.risch [at] deepset.ai`
|
| 54 |
+
- Malte Pietsch: `malte.pietsch [at] deepset.ai`
|
| 55 |
+
- Michel Bartels: `michel.bartels [at] deepset.ai`
|
| 56 |
+
## About us
|
| 57 |
+
|
| 58 |
+
We bring NLP to the industry via open source!
|
| 59 |
+
Our focus: Industry specific language models & large scale QA systems.
|
| 60 |
+
|
| 61 |
+
Some of our work:
|
| 62 |
+
- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
|
| 63 |
+
- [GermanQuAD and GermanDPR datasets and models (aka "gelectra-base-germanquad", "gbert-base-germandpr")](https://deepset.ai/germanquad)
|
| 64 |
+
- [FARM](https://github.com/deepset-ai/FARM)
|
| 65 |
+
- [Haystack](https://github.com/deepset-ai/haystack/)
|
| 66 |
+
|
| 67 |
+
Get in touch:
|
| 68 |
+
[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Slack](https://haystack.deepset.ai/community/join) | [GitHub Discussions](https://github.com/deepset-ai/haystack/discussions) | [Website](https://deepset.ai)
|
| 69 |
+
|
| 70 |
+
By the way: [we're hiring!](http://www.deepset.ai/jobs)
|