metadata
license: mit
datasets:
- philipphager/baidu-ultr-pretrain
- philipphager/baidu-ultr_uva-mlm-ctr
metrics:
- dcg@1
- dcg@3
- dcg@5
- dcg@10
- ndcg@10
- mrr@10
Naive Listwise MonoBERT trained on Baidu-ULTR
A flax-based MonoBERT cross encoder trained on the Baidu-ULTR dataset with a listwise softmax cross-entropy loss on clicks. The loss is called "naive" as we use user clicks as a signal of relevance without any additional position bias correction. For more info, read our paper here.
Usage
from src.model import ListwiseCrossEncoder
model = ListwiseCrossEncoder.from_pretrained(
"philipphager/baidu-ultr_uva-bert_naive-listwise",
)
model(batch)
Test Results on Baidu-ULTR Expert Annotations
| Model | log-likelihood | DCG@1 | DCG@3 | DCG@5 | DCG@10 | nDCG@10 | MRR@10 |
|---|---|---|---|---|---|---|---|
| Naive Pointwise | 0.2272 | 1.6836 | 3.5616 | 4.8822 | 7.4244 | 0.3640 | 0.6096 |
| Naive Listwise | - | 1.9738 | 4.1609 | 5.6861 | 8.5432 | 0.4091 | 0.6436 |