Create README.md

README.md

@@ -0,0 +1,79 @@
+---
+license: mit
+datasets:
+- philipphager/baidu-ultr-pretrain
+- philipphager/baidu-ultr_uva-mlm-ctr
+metrics:
+- log-likelihood
+- dcg@1
+- dcg@3
+- dcg@5
+- dcg@10
+- ndcg@10
+- mrr@10
+---
+
+# Listwise MonoBERT trained on Baidu-ULTR with Inverse Propensity Scoring (IPS)
+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 with IPS correction** adopted based on the work by [Ai et al](https://arxiv.org/abs/1804.05938). The loss uses inverse propensity scoring to mitigate position bias in click data by weighting clicks on items higher that are less likely to be observed by users. 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
+
+
+
+
+## Usage
+Here is an example of downloading the model and calling it for inference on a mock batch of input data. For more details on how to use the model on the Baidu-ULTR dataset, take a look at our [training](https://github.com/philipphager/baidu-bert-model/blob/main/main.py) and [evaluation scripts](https://github.com/philipphager/baidu-bert-model/blob/main/eval.py) in our code repository.
+
+```Python
+import jax.numpy as jnp
+
+from src.model import ListwiseIPSCrossEncoder
+
+model = ListwiseIPSCrossEncoder.from_pretrained(
+    "philipphager/baidu-ultr_uva-bert_ips-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, train=False)
+print(outputs)
+```
+
+## 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},
+}
+```
+