| # Dense Encoder - Distilbert - Frozen Token Embeddings | |
| This model is a distilbert-base-uncased model trained for 30 epochs (235k steps), 64 batch size with MarginMSE Loss on MS MARCO dataset. | |
| The token embeddings were frozen. | |
| | Dataset | Model with updated token embeddings | Model with frozen embeddings | | |
| | --- | :---: | :---: | | |
| | TREC-DL 19 | 70.68 | 68.60 | | |
| | TREC-DL 20 | 67.69 | 70.21 | | |
| | FiQA | 28.89 | 28.60 | | |
| | Robust04 | 39.56 | 39.08 | | |
| | TREC-COVID v2 | 69.80 | 69.84 | | |
| | TREC-NEWS | 37.97 | 38.27 | | |
| | Avg. 4 BEIR tasks | 44.06 | 43.95 | | |