SZLHOLDINGS/miniembed-nano-r1-candidate

MiniEmbed-Nano r1 candidate (PROPOSAL-ONLY, measured candidate)

Status: proposal-only candidate. This is a real, trained sentence-transformers embedder intended to replace the SZLHOLDINGS/MiniEmbed-Nano hash-table fixture (64x12 SHA-256 table, not neural, never a real embedder) for the szl-skills research-memory retrieval role. The fixture repo is NOT overwritten; this is a separate candidate repo. Do not wire this into any skill until a human accepts it.

Measured eval (this run only; see eval_report.json)

Held-out slice of the two training corpora, group-split by source file so no eval chunk shares a source group with any training pair. Weak labels: chunk title as query, chunk text as positive. Binary relevance, shared corpus.

metric	miniembed-r1 candidate	BM25 lexical baseline
nDCG@10	0.608145	0.527441
recall@1	0.528986	0.460145
recall@5	0.644928	0.554348
recall@10	0.702899	0.597826

Eval size: 276 queries over 276 held-out chunks. This is title-query recall on two SZL corpora only — not an MTEB score and not evidence of general retrieval quality.

Provenance

• Base: sentence-transformers/all-MiniLM-L6-v2 @ 1110a243fdf4706b3f48f1d95db1a4f5529b4d41
• Data: SZLHOLDINGS/szl-second-brain-inrepo + SZLHOLDINGS/rag-corpus-v1 (titles -> chunk text, weak labels, seeded split); rejected: frontier-candidates (handles-only) and a11oy-verifiable-corpus (receipt logs)
• Training pairs sha256 (pairs.train.jsonl): 22adee30dadd02111732a89511750219331f60e458dad91852503cceda4a4c0f
• Seed 20260721, 2 epochs, batch 32, lr 2e-05, MultipleNegativesRankingLoss, final train loss 0.843704
• Trained 2026-11-05T02:02:21.837202+00:00 UTC on hf jobs uv run (flavor t4-small recommended)

Honesty

Claim	Label
Neural embedder trained by SGD	TRUE (this candidate)
nDCG@10 / recall@{1,5,10} above	MEASURED (held-out, in-run)
Beats BM25 nDCG@10 on this slice	MEASURED TRUE
General-purpose retrieval quality	UNAVAILABLE
MTEB numbers	NONE PASTED
Overwrites MiniEmbed-Nano	FALSE — separate candidate repo

Energy UNAVAILABLE. Trained by szl-estate/jobs/train_miniembed.py.

The estate

Surface	Link
Company	szlholdings.com
Sponsors	Become a sponsor
GitHub	szl-holdings
Receipts chain	frontier-54-runs
Live dashboard	SI command center