Cross-Encoder for MS Marco
This model uses BERT-Tiny, a tiny BERT model with only 2 layers, 2 attention heads and 128 dimension size.
It was trained on MS Marco Passage Ranking task.
The model can be used for Information Retrieval: Given a query, encode the query will all possible passages (e.g. retrieved with ElasticSearch). Then sort the passages in a decreasing order. See SBERT.net Information Retrieval for more details. The training code is available here: SBERT.net Training MS Marco
Usage and Performance
Pre-trained models can be used like this:
from sentence_transformers import CrossEncoder
model = CrossEncoder('model_name', max_length=512)
scores = model.predict([('Query', 'Paragraph1'), ('Query', 'Paragraph2') , ('Query', 'Paragraph3')])
In the following table, we provide various pre-trained Cross-Encoders together with their performance on the TREC Deep Learning 2019 and the MS Marco Passage Reranking dataset.
| Model-Name | NDCG@10 (TREC DL 19) | MRR@10 (MS Marco Dev) | Docs / Sec (BertTokenizerFast) | Docs / Sec (Python Tokenizer) |
|---|---|---|---|---|
| cross-encoder/ms-marco-TinyBERT-L-2 | 67.43 | 30.15 | 9000 | 780 |
| cross-encoder/ms-marco-TinyBERT-L-4 | 68.09 | 34.50 | 2900 | 760 |
| cross-encoder/ms-marco-TinyBERT-L-6 | 69.57 | 36.13 | 680 | 660 |
| cross-encoder/ms-marco-electra-base | 71.99 | 36.41 | 340 | 340 |
| Other models | ||||
| nboost/pt-tinybert-msmarco | 63.63 | 28.80 | 2900 | 760 |
| nboost/pt-bert-base-uncased-msmarco | 70.94 | 34.75 | 340 | 340 |
| nboost/pt-bert-large-msmarco | 73.36 | 36.48 | 100 | 100 |
| Capreolus/electra-base-msmarco | 71.23 | 340 | 340 | |
| amberoad/bert-multilingual-passage-reranking-msmarco | 68.40 | 330 | 330 |
Note: Runtime was computed on a V100 GPU. A bottleneck for smaller models is the standard Python tokenizer from Huggingface in version 3. Replacing it with the fast tokenizer based on Rust, the throughput is significantly improved: