Model
BERT Distilled 4L-288D PyTorch checkpoint distilled on SQuAD v1.1 dataset using AMP
Use the NGC CLI to download:
Copied!
1 Version
21.11.0_ampSelected07/07/2022 2:25 PM UTC49.6 MBAccuracy: 010 EpochsBatch Size: 256GPU: V100 Copied!
21.11.0_ampSelected
07/07/2022 2:25 PM UTC49.6 MBAccuracy: 010 EpochsBatch Size: 256GPU: V100
Copied!
architecture
| Key | Value |
|---|---|
| attn_heads | 12 |
| task | SQuAD v1.1 |
| hidden_size | 288 |
| layers | 4 |
| type | Dist-4L-288D |
performance
| Key | Value |
|---|---|
| F1 | 83.09 |
training
| Key | Value |
|---|---|
| batch_size | 256 |
| LR | 3e-05 |
| training_precision | AMP |
| max_seq_length | 384 |
| epochs | 10 |
| dataset | SQuAD v1.1 |
| warmup_proportion | 0.1 |