BoDong's picture
First model version
4d03cd5
|
raw
history blame
769 Bytes
metadata
license: mit

Sparse BERT mini model (uncased)

Finetuned model pruned to 1:4 structured sparsity. The model is a pruned version of the BERT mini model.

Intended Use

The model can be used for inference with sparsity optimisztion. For further details on the model and its usage, see our repo and our implementation available here. We also upload the quanted int8 BERT mini sparse Neural Engine IR (acc 87.15) here, could be directly used by NLP Toolkit ref inference.

Evaluation Results

We get the following results on the sst2 tasks development set:

Task SST-2 (Acc)
87.2