Update README.md

README.md

@@ -1,3 +1,64 @@
----
-license: gemma
----
+---
+pipeline_tag: text-classification
+tags:
+- transformers
+- information-retrieval
+language: pl
+license: gemma
+
+---
+
+<h1 align="center">polish-reranker-roberta-v2</h1>
+
+This is an improved version of reranker based on [sdadas/polish-roberta-large-v2](https://huggingface.co/sdadas/polish-roberta-large-v2) trained with [RankNet loss](https://icml.cc/Conferences/2015/wp-content/uploads/2015/06/icml_ranking.pdf) on a large dataset of text pairs.
+The model was trained in the same way and on the same data as [sdadas/polish-roberta-large-ranknet](https://huggingface.co/sdadas/polish-roberta-large-ranknet), with the following improvements:
+- We used predictions from [BAAI/bge-reranker-v2.5-gemma2-lightweight](https://huggingface.co/BAAI/bge-reranker-v2.5-gemma2-lightweight) for distillation instead of [unicamp-dl/mt5-13b-mmarco-100k](https://huggingface.co/unicamp-dl/mt5-13b-mmarco-100k).
+- We used a custom implementation of the RoBERTa with support for Flash Attention 2. If you want to use these features, load the model with the arguments `trust_remote_code=True` and `attn_implementation="flash_attention_2"`.
+
+Our reranker achieves results close to [BAAI/bge-reranker-v2.5-gemma2-lightweight](https://huggingface.co/BAAI/bge-reranker-v2.5-gemma2-lightweight) on the PIRB benchmark, even outperforming it on some datasets. At the same time, it is over 21 times smaller — 435M vs. 9.24B parameters.
+
+## Usage (Huggingface Transformers)
+
+The model can be used with Huggingface Transformers in the following way:
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import numpy as np
+
+query = "Jak dożyć 100 lat?"
+answers = [
+    "Trzeba zdrowo się odżywiać i uprawiać sport.",
+    "Trzeba pić alkohol, imprezować i jeździć szybkimi autami.",
+    "Gdy trwała kampania politycy zapewniali, że rozprawią się z zakazem niedzielnego handlu."
+]
+
+model_name = "sdadas/polish-reranker-roberta-v2"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(
+    model_name, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
+)
+texts = [f"{query}</s></s>{answer}" for answer in answers]
+tokens = tokenizer(texts, padding="longest", max_length=512, truncation=True, return_tensors="pt")
+output = model(**tokens)
+results = output.logits.detach().numpy()
+results = np.squeeze(results)
+print(results.tolist())
+```
+
+## Evaluation Results
+
+The model achieves **NDCG@10** of **65.30** in the Rerankers category of the Polish Information Retrieval Benchmark. See [PIRB Leaderboard](https://huggingface.co/spaces/sdadas/pirb) for detailed results.
+
+## Citation
+
+```bibtex
+@article{dadas2024assessing,
+  title={Assessing generalization capability of text ranking models in Polish}, 
+  author={Sławomir Dadas and Małgorzata Grębowiec},
+  year={2024},
+  eprint={2402.14318},
+  archivePrefix={arXiv},
+  primaryClass={cs.CL}
+}
+```