Skip to content

k-nlp/klej-submission

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 

KLEJ submissions

Details of KLEJ submissions: https://klejbenchmark.com/leaderboard/

XLM-RoBERTa (large)

Library: transformers Model: xlm-roberta-large

  • effective batch size: 32
  • learning rate: 2e-5
  • adam epsilon: 1e-8
  • num train epochs: 4
  • weight decay: 0.0
  • max grad norm: 1.0
  • warmup steps: 100
Run NKJP-NER CDSC-E CDSC-R CBD PolEmo2.0-IN PolEmo2.0-OUT DYK PSC AR Avg
1 94.6 94.4 94.7 50.7 90.4 79.8 71.6 98.2 87.5 84.7

XLM-RoBERTa (large) - FT2

Based on submission: XLM-RoBERTa (large)

The only change is for the CBD task:

  • every positive example is duplicated 4 times
  • num train epochs: 1
  • weight decay: 0.01
  • max grad norm: 5.0
  • learning rate: 1e-5
Run NKJP-NER CDSC-E CDSC-R CBD PolEmo2.0-IN PolEmo2.0-OUT DYK PSC AR Avg
1 94.6 94.4 94.7 67.9 90.4 79.8 71.6 98.2 87.5 86.6

XLM-RoBERTa (large) - FT3

Library: fairseq Model: xlm-roberta-large

Models were trained using original scripts from Polish RoBERTa.

Run NKJP-NER CDSC-E CDSC-R CBD PolEmo2.0-IN PolEmo2.0-OUT DYK PSC AR Avg
1 94,1 94,4 94,7 70,6 92,4 81 72,8 98,9 88,4 87,5
2 94,7 94,3 94,6 67,4 92,5 81,6 73,9 98,0 89,3 87,4
3 94,6 94,5 95 67,2 92,4 82 73,4 98,8 89,1 87,4
4 94,8 94,3 94,8 67,5 92,7 80,6 75,3 98,3 88,5 87,4
5 94,4 94,0 94,8 68,5 92,5 81,2 77,1 98,6 88,5 87,7

XLMR+NKJP

XLM-RoBERTa (large) model has been further trained on NCP (National Corpus of Polish).

Models were fine-tuned using original scripts from Polish RoBERTa.

Run NKJP-NER CDSC-E CDSC-R CBD PolEmo2.0-IN PolEmo2.0-OUT DYK PSC AR Avg
1 94,2 94,2 94,5 72,4 93,1 77,9 77,5 98,6 88,2 87,8
2 94,9 94,5 94,6 68,2 92,8 82 74,6 99,1 88,5 87,7
3 95,1 94,5 94,7 67,8 91,8 81,8 75 97,9 88,2 87,4
4 94,7 93,5 94,8 69,2 92,2 82,8 73,6 98 88,7 87,5
5 95,2 94,4 94,7 70,3 92,7 83,8 76,6 98,9 88,8 88,4
Avg 94,8 94,2 94,7 69,6 92,5 81,7 75,5 98,5 88,5 87,8

Contact

LinkedIn: https://www.linkedin.com/in/wrobelkrzysztof/

About

Details of KLEJ submissions.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published