If you trained a Sentence Transformers model before v3.0, everything you know about the fitting process is now legacy. The SentenceTransformer.fit API has been officially retired and replaced by the SentenceTransformerTrainer, and the old tutorial now carries a deprecation notice at the top pointing users to four new guides covering embedding models, rerankers, sparse embeddings, and multimodal architectures.
The architectural fundamentals haven’t changed. A Sentence Transformer still runs text through a pretrained Transformer backbone — the original guide used distilroberta-base — then pools the contextualized token embeddings into a single fixed-size vector. Mean pooling is the default. And the case for this architecture over raw BERT remains as stark as ever: finding the most similar pair among 10,000 sentences takes roughly 65 hours with vanilla BERT versus about 5 seconds with a Sentence Transformer. BERT with averaged token embeddings even underperforms GloVe vectors from 2014, which tells you how far sentence-level representation has come since then.
What the split into four separate guides signals is that the training problem has fragmented. Embedding models, cross-encoder rerankers, sparse retrieval models, and multimodal systems each have different data formats and loss functions. A single monolithic tutorial can’t do justice to all of them anymore. The new Trainer API brings Sentence Transformers in line with the Hugging Face ecosystem’s standard training workflow, the same shift Transformers made when it moved from custom training loops to the Trainer class years ago.
For practitioners mid-project on the old API, this is a breaking change with real migration cost. For everyone else, it’s a sign the library has matured enough to warrant task-specific documentation. The historical content remains available below the deprecation notice, but relying on it for new work is asking for compatibility headaches.
