Update README.md
Browse files
README.md
CHANGED
@@ -13,6 +13,8 @@ The model can be used for fine-tuning to downstream tasks with sparsity already
|
|
13 |
To keep the sparsity a mask should be added to each sparse weight blocking the optimizer from updating the zeros.
|
14 |
|
15 |
## Evaluation Results
|
|
|
|
|
16 |
| Task | MNLI-m (Acc) | MNLI-mm (Acc) | QQP (Acc/F1) | QNLI (Acc) | SST-2 (Acc) | STS-B (Pears/Spear) | SQuADv1.1 (Acc/F1) |
|
17 |
|------|--------------|---------------|--------------|------------|-------------|---------------------|--------------------|
|
18 |
| | 83.3 | 83.9 | 90.8/87.6 | 90.4 | 91.3 | 88.8/88.3 | 80.5/88.2 |
|
|
|
13 |
To keep the sparsity a mask should be added to each sparse weight blocking the optimizer from updating the zeros.
|
14 |
|
15 |
## Evaluation Results
|
16 |
+
We get the following results on the tasks development set, all results are mean of 5 different seeded models:
|
17 |
+
|
18 |
| Task | MNLI-m (Acc) | MNLI-mm (Acc) | QQP (Acc/F1) | QNLI (Acc) | SST-2 (Acc) | STS-B (Pears/Spear) | SQuADv1.1 (Acc/F1) |
|
19 |
|------|--------------|---------------|--------------|------------|-------------|---------------------|--------------------|
|
20 |
| | 83.3 | 83.9 | 90.8/87.6 | 90.4 | 91.3 | 88.8/88.3 | 80.5/88.2 |
|