Training and Finetuning Reranker Models with Sentence Transformers

Sentence Transformers is a Python library for using and training embedding and reranker models for a wide range of applications, such as retrieval augmented generation, semantic search, semantic textual similarity, paraphrase mining, and more. In this blogpost, I'll show you how to use it to finetune a reranker model (also known as a cross-encoder) that beats all existing options on exactly your data. This method can also train extremely strong new reranker models from scratch.

Finetuning reranker models involves several components: datasets, loss functions, training arguments, evaluators, and the trainer class itself. I'll have a look at each of these components, accompanied by practical examples of how they can be used for finetuning strong reranker models.

Lastly, in the Evaluation section, I'll show you that my small finetuned tomaarsen/reranker-ModernBERT-base-gooaq-bce reranker model that I trained alongside this blogpost easily outperforms the 13 most commonly used public reranker models on my evaluation dataset. It even beats models that are 4x bigger.

However, this strength comes with a trade-off: Cross Encoder models are slower as they process every possible pair of texts (e.g., 10 queries with 500 candidate documents requires 5,000 computations instead of 510 for embedding models). This makes them less efficient for large-scale initial retrieval but ideal for reranking: refining the top-k results first identified by faster Sentence Transformer models. The strongest search systems commonly use this 2-stage "retrieve and rerank" approach.

You can customize the training process using the CrossEncoderTrainingArguments class. This class lets you adjust parameters that can impact training speed and help you understand what's happening during training.

The mine_hard_negatives function has a convenient include_positives parameter, which can be set to True to also mine for the positive texts. When supplied as documents (which have to be 1. ranked and 2. contain positives) to CrossEncoderRerankingEvaluator , the evaluator will not just evaluate the reranking performance of the CrossEncoder, but also the original rankings by the embedding model used for mining.

I use the BinaryCrossEntropyLoss , which is well suited for these labeled pairs. I also set up 2 forms of evaluation: CrossEncoderNanoBEIREvaluator which evaluates against the NanoBEIR benchmark and CrossEncoderRerankingEvaluator which evaluates the performance of reranking the top 30 results from the aforementioned static embedding model. Afterwards, I define a fairly standard set of hyperparameters, including learning rates, warmup ratios, bf16, loading the best model at the end, and some debugging parameters. Lastly, I run the trainer, perform post-training evaluation, and save the model both locally and on the Hugging Face Hub.

The CrossEncoderTrainer uses datasets.Dataset or datasets.DatasetDict instances for training and evaluation. You can load data from the Hugging Face Datasets Hub or use your local data in whatever format you prefer (e.g. CSV, JSON, Parquet, Arrow, or SQL).

It is important that your dataset format matches your loss function (or that you choose a loss function that matches your dataset format and model). Verifying whether a dataset format and model work with a loss function involves three steps:

Evaluation results are automatically stored in the generated model card upon saving a model, alongside the base model, language, license, evaluation results, training & evaluation dataset info, hyperparameters, training logs, and more. Without any effort, your uploaded models should contain all the information that your potential users would need to determine whether your model is suitable for them.

Repeating the recipe with a bigger base model results in tomaarsen/reranker-ModernBERT-large-gooaq-bce , a reranker model that blows all existing general-purpose reranker models out of the water on my data.

If you're interested in finetuning embedding models instead, then consider reading through my prior Training and Finetuning Embedding Models with Sentence Transformers blogpost as well.

Reranker models, often implemented using Cross Encoder architectures, are designed to evaluate the relevance between pairs of texts (e.g., a query and a document, or two sentences). Unlike Sentence Transformers (a.k.a. bi-encoders, embedding models), which independently embed each text into vectors and compute similarity via a distance metric, Cross Encoder process the paired texts together through a shared neural network, resulting in one output score. By letting the two texts attend to each other, Cross Encoder models can outperform embedding models.