Model mirror: all-MiniLM-L6-v2

Not a software release. An unmodified mirror of sentence-transformers/all-MiniLM-L6-v2, Apache-2.0, for lobslaw's built-in embedder.

Why mirror it. The embedder downloads a model on first boot. Fetching from a release rather than HuggingFace removes a third party from that path: the bytes are pinned to this tag, cannot change or be withdrawn under a running deployment, and CI does not depend on another host's availability or rate limits.

Use it:

[compute.embeddings]
type         = "builtin"
model        = "all-MiniLM-L6-v2"
download_url = "https://github.com/jmylchreest/lobslaw/releases/download/models-all-MiniLM-L6-v2"

What changed. Nothing in the weights or tokenizer — config.json, model.safetensors and tokenizer.json are byte-identical to upstream (see SHA256SUMS). One rename: 1_Pooling/config.json becomes 1_Pooling.config.json, because a release asset filename cannot contain /. lobslaw's fetcher restores the nested path, so what lands on disk is an ordinary HuggingFace snapshot.

91 MB. English only. See NOTICE for attribution and LICENSE for the Apache-2.0 text.

The tag deliberately does not match v*, so this does not trigger the container release workflow.