We utilised RankNet loss and trained the model on the same data as sdadas/polish-reranker-roberta-v2
BAAI/bge-reranker-v2.5-gemma2-lightweight was used as the teacher model for distillation
After the training, we merged the original and fine-tuned weights to create the final checkpoint
We used a custom implementation of XLM-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". This is especially important for this model, since BAAI/bge-reranker-v2-m3 supports long contexts of 8192 tokens. For such input length, the inference can be up to 400% faster with Flash Attention in comparison to the original model.

In most cases, the use of sdadas/polish-reranker-roberta-v2 is preferred to this model as it achieves better results for Polish. The main advantage of this model is its context length, so it may perform better on some datasets with long documents.

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-bge-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",
    device_map="cuda"
)
texts = [f"{query}</s></s>{answer}" for answer in answers]
tokens = tokenizer(texts, padding="longest", max_length=8192, truncation=True, return_tensors="pt").to("cuda")
output = model(**tokens)
results = output.logits.detach().cpu().float().numpy()
results = np.squeeze(results)
print(results.tolist())

Evaluation Results

The model achieves NDCG@10 of 64.21 in the Rerankers category of the Polish Information Retrieval Benchmark. See PIRB Leaderboard 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}
}