# Would have been 35th place 🥲 [LB: 0.75 | PB: 0.745]

:(
I was kind of expecting the shakeup because of the saturated leaderboard and no other models were released with a higher LB.
But to my surprise, MuRIL-Large outperformed xlm-roberta-large (Not in the LB). The xlm-roberta was overfitted to the public LB.
Muril-Large was introduced lately to the competitors and thanks to
[@simrankhanuja](https://www.kaggle.com/simrankhanuja)
for introducing the model.
The Solution :
Simple blend of 5-fold of MuRIL-Large trained on chaii, mlqa and xquad dataset.
The public LB score was 0.75, did some tweaking, the score decreased.
Checked the CV, then lost tracks :)
No-one retrained the MuRIL model so no 0.745 score (my trained model was private, now made public)
This is the
would have been
35th place solution:
[MuRIL-Large | mlqa, xquad, chaii | INFER 💫](https://www.kaggle.com/kishalmandal/muril-large-mlqa-xquad-chaii-infer)
The treasure was in-front my eyes, just that it was covered with mud :(
And as
[@cdeotte](https://www.kaggle.com/cdeotte)
says,
'Trust your CV'
Since this was my first competition I will keep that in mind.
Anyways these were my kernels and all were public :)
xlm-RoBERTa:
[5 folds | INFER | combined model (0.792)](https://www.kaggle.com/kishalmandal/5-folds-infer-combined-model-0-792)
: The shakeup model
[chaii | FIT - 2 epochs | mlqa, xquad, chaii](https://www.kaggle.com/kishalmandal/chaii-fit-2-epochs-mlqa-xquad-chaii)
: Training kernel 1
[chaii | FIT - 7 epochs | Extra Tamil Data](https://www.kaggle.com/kishalmandal/chaii-fit-7-epochs-extra-tamil-data)
: Training kernel 2
MuRIL-Large:
[MuRIL-Large | mlqa, xquad, chaii | FIT 💫](https://www.kaggle.com/kishalmandal/muril-large-mlqa-xquad-chaii-fit)
: Training kernel
[MuRIL-Large | mlqa, xquad, chaii | INFER 💫](https://www.kaggle.com/kishalmandal/muril-large-mlqa-xquad-chaii-infer)
: Inference kernel
EDA of CV scores of different models (xlm-rob):
[EDA + Experimental results 💫](https://www.kaggle.com/kishalmandal/eda-experimental-results)