Hugging Face's stsb-bert-tiny Packs a Full Embedding Stack Into 17.6 MB
A 4.39M-parameter BERT sentence encoder producing 128-dimensional embeddings sits on Hugging Face as a lightweight fixture for testing sentence-transformers pipelines.
- Hugging Face hosts stsb-bert-tiny-safetensors, a 4.39M-parameter BERT sentence encoder
- Outputs 128-dim embeddings via mean pooling, max sequence length of 512 tokens
- Trained 10 epochs with CosineSimilarityLoss, AdamW at 8e-5, batch size 16
- Ships PyTorch and safetensors weights, works with sentence-transformers and raw Transformers
- Racked up 1.84M downloads last month as a CI and prototyping fixture
- Not intended for production retrieval; swap in MiniLM or BGE for real workloads
A tiny BERT checkpoint for full-stack embedding tests
Hugging Face hosts stsb-bert-tiny-safetensors, a 4.39 million-parameter sentence encoder published by the sentence-transformers-testing organization. It converts sentences and paragraphs into 128-dimensional vectors for semantic search, clustering, similarity scoring, and pipeline tests.
The repository includes SafeTensors weights, a PyTorch checkpoint, tokenizer files, pooling configuration, and examples for both Sentence Transformers and raw Transformers. That complete package lets developers exercise the same loading, tokenization, inference, pooling, and similarity paths used by larger embedding models while reducing download size, memory use, and CPU work.
Small weights, complete stack
| Property | Value |
|---|---|
| Parameters | 4.39 million |
| Weight precision | 32-bit floating point |
| Raw parameter storage | About 17.6 MB |
| Embedding size | 128 dimensions |
| Maximum sequence length | 512 tokens |
| Architecture | BertModel followed by mean-token pooling |
| Weight formats | SafeTensors and PyTorch |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
