cstr/splade-v3-GGUF

SPLADE-v3 — GGUF (CrispEmbed)

GGUF conversions of naver/splade-v3 for sparse (learned-lexical) retrieval with CrispEmbed — a dependency-free C/C++ embedding runtime (--sparse mode / crispembed_encode_sparse).

SPLADE produces a sparse vector of weighted vocabulary term expansions (log(1 + ReLU(MLM_logits)), max-pooled over tokens), not a dense embedding.

Files

File	Quant	Size	Sparse-cos vs HF fp32
splade-v3-iq4_xs.gguf	IQ4_XS + imatrix	68 MB	0.9971 (compact default)
splade-v3-q8_0.gguf	Q8_0	111 MB	1.0000
splade-v3-f16.gguf	F16	256 MB	1.0000 (precision control)

Sparse-cos = cosine over the full 30522-dim vocabulary term-weight vector vs the original PyTorch naver/splade-v3 (BertForMaskedLM + SPLADE pooling), averaged over a probe set. The IQ4_XS build uses a CrispEmbed importance matrix collected over a calibration corpus. The MLM/SPLADE head is preserved and verified present in every quant.

Usage

crispembed -m splade-v3-iq4_xs.gguf --sparse "your query text"
# → token_id  weight   (one per expansion term)

License & attribution

Derived from naver/splade-v3, licensed CC-BY-NC-SA-4.0 (non-commercial, share-alike). These GGUF conversions inherit the same license: non-commercial use only, attribution to Naver required, and derivatives must be shared alike. See the base model card and the SPLADE papers for citation.