Wyrdsekai embedding models (ONNX, int8)
Int8 ONNX quantizations of open sentence-transformer encoders, repackaged so wyrdsekai installers have a stable, versioned mirror. These are not wyrdsekai fine-tunes (with one noted exception) — credit and licenses belong to the upstream authors.
| File | Upstream | Role in wyrdsekai |
|---|---|---|
paraphrase-multilingual-MiniLM-L12-v2-q8.onnx (+tokenizer, paraphrase-l12/) |
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | Default retrieval encoder (384-d): memory, library and tool-description search. Installed by wyrd setup. |
paraphrase-multilingual-MiniLM-L12-v2-setfit-q8.onnx |
same base, SetFit-tuned by wyrdsekai | Classifier-head encoder (task routing). Tuned on synthetic labelled corpora only; kept separate from the retrieval encoder on purpose — SetFit tuning degrades retrieval quality. |
minilm-l6-v2-q8.onnx (+tokenizer) |
sentence-transformers/all-MiniLM-L6-v2 | Lightweight fallback for constrained devices. |
All upstream models are Apache 2.0; the quantizations carry the same license.
Hardware permitting, wyrd embedding-model download bge-m3 upgrades retrieval to
BAAI/bge-m3 (fetched from its own repo, not
mirrored here).