rust-sbert

For distiluse-base-multilingual-cased the supported languages are: Arabic, Chinese, Dutch, English, French, German, Italian, Korean, Polish, Portuguese, Russian, Spanish, Turkish. Performance on the extended STS2017: 80.1.

Usage

Example

The API is made to be very easy to use and enables you to create quality multilingual sentence embeddings in a straightforward way.

Load a model by pointing at its directory (must contain modules.json and the subdirectories/paths it references):

let mut home: PathBuf = env::current_dir().unwrap();
home.push("path-to-model");

You can use different versions of the models that use different tokenizers:

// To use Hugging Face tokenizer
let sbert_model = SBertHF::new(home, None).unwrap();

// To use Rust-tokenizers
let sbert_model = SBertRT::new(home, None).unwrap();

SBert<T> is a backward-compat alias for the underlying SentenceTransformer<T> struct — both names refer to the same type, so existing call sites keep working unchanged.

Now, you can encode your sentences:

let texts = vec![
    "You can encode".to_string(),
    "As many sentences".to_string(),
    "As you want".to_string(),
    "Enjoy ;)".to_string(),
];

let batch_size = 64;

let output = sbert_model.forward(&texts, batch_size).unwrap();

The parameter batch_size can be left to None to let the model use its default value.

Then you can use the output sentence embedding in any application you want.

Loading a bert-backbone checkpoint (e.g. all-MiniLM-L6-v2)

rust-bert publishes pre-converted rust_model.ot weights for a handful of well-known checkpoints (those its own sentence-embeddings pipeline lists — all-MiniLM-L6-v2 is one of them), so no Python toolchain is needed for those. For anything else you must convert the weights yourself (see the next section); note that hub checkpoints shipping 2_Dense/model.safetensors are not covered by utils/prepare_models.py either. The HF modules.json for all-MiniLM-L6-v2 declares the Transformer at path: "" (files at the model root):

mkdir -p models/all-MiniLM-L6-v2/1_Pooling
HF=https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main

curl -L -o models/all-MiniLM-L6-v2/model.ot                $HF/rust_model.ot
curl -L -o models/all-MiniLM-L6-v2/config.json             $HF/config.json
curl -L -o models/all-MiniLM-L6-v2/vocab.txt               $HF/vocab.txt
curl -L -o models/all-MiniLM-L6-v2/tokenizer_config.json   $HF/tokenizer_config.json
curl -L -o models/all-MiniLM-L6-v2/sentence_bert_config.json $HF/sentence_bert_config.json
curl -L -o models/all-MiniLM-L6-v2/special_tokens_map.json $HF/special_tokens_map.json
curl -L -o models/all-MiniLM-L6-v2/1_Pooling/config.json   $HF/1_Pooling/config.json
curl -L -o models/all-MiniLM-L6-v2/modules.json            $HF/modules.json

Then load as usual: SBertRT::new("models/all-MiniLM-L6-v2", None).

Convert distilbert models from Python to Rust

For checkpoints that don't ship a pre-converted .ot (e.g. the original distiluse-base-multilingual-cased export), convert the PyTorch weights manually.

Firstly, get a model provided by UKPLabs (all models are here):

mkdir -p models/distiluse-base-multilingual-cased

wget -P models https://public.ukp.informatik.tu-darmstadt.de/reimers/sentence-transformers/v0.2/distiluse-base-multilingual-cased.zip

unzip models/distiluse-base-multilingual-cased.zip -d models/distiluse-base-multilingual-cased

Then, you need to convert the model in a suitable format (requires pytorch):

python utils/prepare_distilbert.py models/distiluse-base-multilingual-cased

A dockerized environment is also available for running the conversion script:

docker build -t tch-converter -f utils/Dockerfile .

docker run \
  -v $(pwd)/models/distiluse-base-multilingual-cased:/model \
  tch-converter:latest \
  python prepare_distilbert.py /model

Finally, set "output_attentions": true in distiluse-base-multilingual-cased/0_distilbert/config.json.

Pipeline architecture

SentenceTransformer<T> holds:

• transformer: Box<dyn TransformerBackend> — dispatches on config.json's model_type field. Implemented: DistilBertBackend, BertBackend.
• post: Vec<Box<dyn Module>> — built from modules.json in declared order. The forward pass threads a Features enum (Token { .. } → Sentence { .. }) through each module; Pooling does the token→sentence reduction.

To extend:

• New backbone (e.g. RoBERTa): add a RobertaBackend struct + impl TransformerBackend for it + add a match arm in modules::transformer::load. Note this alone is not enough: tokenization is hard-coded to vocab.txt + WordPiece in SentenceTransformer::new, so a backbone with a different tokenizer (RoBERTa's SentencePiece/BPE) also needs a new tokenizers::Tokenizer implementation and corresponding tokenizer-file resolution.
• New post-transformer module (e.g. WeightedLayerPooling): add a module struct + impl Module for it + add a match arm on short_type in SentenceTransformer::new.

See src/modules/ for the building blocks.