README.md

---
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](https://arxiv.org/abs/2207.03051) 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](https://arxiv.org/abs/2404.02543) and [find the code for this model here](https://github.com/philipphager/baidu-bert-model).

## 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 |

## Usage
Here is an example with a mock input batch for how to download and call the model:

```Python
import jax.numpy as jnp

from src.model import ListwiseCrossEncoder

model = ListwiseCrossEncoder.from_pretrained(
    "philipphager/baidu-ultr_uva-bert_naive-listwise",
)

# Mock batch following Baidu-ULTR with 4 documents, each with 8 tokens
batch = {
    # Query_id for each document
    "query_id": jnp.array([1, 1, 1, 1]),
    # Document position in SERP
    "positions": jnp.array([1, 2, 3, 4]),
    # Token ids for: [CLS] Query [SEP] Document
    "tokens": jnp.array([
        [2, 21448, 21874, 21436, 1, 20206, 4012, 2860],
        [2, 21448, 21874, 21436, 1, 16794, 4522, 2082],
        [2, 21448, 21874, 21436, 1, 20206, 10082, 9773],
        [2, 21448, 21874, 21436, 1, 2618, 8520, 2860],
    ]),
    # Specify if a token id belongs to the query (0) or document (1)
    "token_types": jnp.array([
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
    ]),
    # Marks if a token should be attended to (True) or ignored, e.g., padding tokens (False):
    "attention_mask": jnp.array([
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
    ]),
}

outputs = model(batch)
print(outputs)
```

For more details on how to use the model with real data from Baidu-ULTR, take a look at the [evaluation script of our model repository](https://github.com/philipphager/baidu-bert-model/blob/main/eval.py).

## Reference
```
@inproceedings{Hager2024BaiduULTR,
  author = {Philipp Hager and Romain Deffayet and Jean-Michel Renders and Onno Zoeter and Maarten de Rijke},
  title = {Unbiased Learning to Rank Meets Reality: Lessons from Baidu’s Large-Scale Search Dataset},
  booktitle = {Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR`24)},
  organization = {ACM},
  year = {2024},
}
```