opensearch-neural-sparse-encoding-multilingual-v1-en — English-pruned SPLADE for Statelet

English-pruned ONNX build of opensearch-project/opensearch-neural-sparse-encoding-multilingual-v1, packaged for the learned-sparse retrieval path in Statelet.

This is an English-only derivative. Do not use it for multilingual sparse retrieval.

What was changed

For a SPLADE encoder the pruning is deeper than an input-table prune, because the output is the MLM-head logit vector [B, T, vocab] — the vocab axis is the sparse representation. Every tensor carrying a vocab-sized axis is sliced:

• input word-embedding rows [V, H] → [K, H]
• MLM decoder projection (tied or not) [H, V] → [H, K]
• decoder bias [V] → [K]
• per-channel quantization scales / zero-points sized [V]
• int64 shape constants containing V (Reshape targets)

plus vocab-sized dims patched in graph outputs and value_info.

Verified end-to-end: pruned SPLADE vectors (keyed by token string) match the original's on English text by cosine over the token union, the only permitted difference being dropped non-Latin tokens.

Index compatibility

Statelet converts logit indices to token strings before writing postings, so the remapped ids are invisible to existing index data — English corpora need no re-ingest when moving from the multilingual model to this one.

Files

file	size	notes
model.onnx	~676 MiB	fp32, default
model_fp16.onnx	~338 MiB	fp16
tokenizer.json	—	pruned tokenizer — must be used with these graphs

An INT8 variant exists but is not published here yet.

Use in Statelet

STATELET_MODEL_ROOT=/absolute/path/to/models

with this repository at $STATELET_MODEL_ROOT/opensearch-neural-sparse-encoding-multilingual-v1-en, or set STATELET_SPARSE_MODEL to point straight at it.

License

Apache-2.0, inherited from the upstream model. See the upstream card for authoritative terms.