3rd-Place-Solution for Feedback-Prize---English-Language-Learning.
Competition Link
The solution write-up Link
order selected by Hill Climbing | backbone | cv score |
---|---|---|
1 | deberta-v3-large | 0.447 |
2 | deberta-v3-large-squad2 | 0.4524 |
3 | deberta-v3-large | 0.4498 |
4 | deberta-large-mnli | 0.4548 |
5 | deberta-v3-large | 0.4492 |
6 | xlm-roberta-large | 0.4575 |
7 | deberta-v3-large-squad2 | 0.457 |
8 | deberta-v3-large-squad2 | 0.4525 |
9 | deberta-v3-large | 0.4489 |
10 | deberta-v3-large-squad2 | 0.4565 |
11 | RAPIDS-SVR | 0.4526 |
12 | deberta-v3-large | 0.4495 |
13 | deberta-v3-large | 0.4502 |
14 | TF-deberta-v3-base | 0.4554 |
15 | deberta-v3-large | 0.4527 |
16 | deberta-v3-large | 0.4522 |
17 | deberta-v3-large | 0.45 |
18 | deberta-v3-large | 0.4516 |
19 | deberta-v3-large | 0.4509 |
20 | deberta-v3-base | 0.4575 |
21 | deberta-v3-large-squad2 | 0.4501 |
22 | deberta-v3-large | 0.4512 |
23 | roberta-large | 0.4571 |
24 | deberta-v3-large | 0.45 |
Build docker image from Dockerfile
In order to run the code, you will need to download the competition data and extract it in the data folder.
To run experiments :
For non deberta models :
- you should change
model_name
inconfigs/non_deberta_config.yaml
- run
python train.py --config configs/non_deberta_config.yaml
For deberta models :
- you should change
model_name
inconfigs/deberta_v3_large.yaml
- run
python train.py --config configs/deberta_v3_large.yaml
For 2xPooling models:
- you should use one of
configs/deberta-v3-large-2xpooling-paragraph.yaml
,configs/deberta-v3-large-2xpooling-sentences.yaml
,configs/deberta-v3-large-2xpooling-words.yaml
- run
python train.py --config configs/deberta-v3-large-2xpooling-paragraph.yaml
For non PyTorch models :
- run rapids-svr-cv-0-450.ipynb in
src/model_zoo folder
- run tf-deberta-v3-base-cv-0-455.ipynb in
src/model_zoo folder
If you wish to only train PyTorch models, skip these two and set NUM_MODELS = 22
in inference code.
To reproduce our final score, run this code from kaggle kernels.
Models were trained using ZbyHP Z8 workstation with Ubuntu 20.04.1 LTS.