MiniEmbed-Nano card
Browse files
README.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: numpy
|
| 4 |
+
tags:
|
| 5 |
+
- governed-ai
|
| 6 |
+
- khipu
|
| 7 |
+
- szl-holdings
|
| 8 |
+
- embedding
|
| 9 |
+
- silhouette
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# MiniEmbed-Nano
|
| 13 |
+
|
| 14 |
+
Tiny hash+table embed: **V=64, d=12**, L2-normalized rows. **Not a foundation embed. Not neural. Not MiniEmbed 3290×128.**
|
| 15 |
+
|
| 16 |
+
Canonical source: [szl-holdings/szl-khipu](https://github.com/szl-holdings/szl-khipu)
|
| 17 |
+
Sibling card: [SZLHOLDINGS/szl-khipu](https://huggingface.co/SZLHOLDINGS/szl-khipu)
|
| 18 |
+
The larger statistical MiniEmbed (3290 × 128) lives on [SZLHOLDINGS/szl-kernels](https://huggingface.co/SZLHOLDINGS/szl-kernels) — a different artifact. Do not mix them.
|
| 19 |
+
|
| 20 |
+
```python
|
| 21 |
+
from szl_khipu.train import mini_embed
|
| 22 |
+
|
| 23 |
+
emb = mini_embed.build(seed=20260721)
|
| 24 |
+
vec = emb.embed("knot the run")
|
| 25 |
+
print(emb.V, emb.D, vec.shape)
|
| 26 |
+
# 64 12 (12,)
|
| 27 |
+
emb.save_npz("mini_embed.npz")
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
## What it does
|
| 31 |
+
|
| 32 |
+
- SHA-256 token id modulo 64. Mean-pool then L2. Deterministic given seed.
|
| 33 |
+
- Built here on CPU NumPy. Honesty **REPORTED**. Energy **UNAVAILABLE**.
|
| 34 |
+
- No analogy score. No retrieval score. No SVD variance claim (that belongs to the 3290×128 table).
|
| 35 |
+
|
| 36 |
+
## What it is NOT
|
| 37 |
+
|
| 38 |
+
- **Not** the [SZLHOLDINGS/szl-kernels](https://huggingface.co/SZLHOLDINGS/szl-kernels) MiniEmbed (3290 × 128, SVD var 0.3146).
|
| 39 |
+
- **Not neural. Not word2vec. Not a foundation embed.**
|
| 40 |
+
- **Not 1.5B. Not Qwen.**
|
| 41 |
+
- **Not proven trust.** Λ uniqueness remains Conjecture 1 OPEN.
|
| 42 |
+
- Energy **UNAVAILABLE**. CUDA **UNAVAILABLE**. Never a fabricated joule.
|
| 43 |
+
|
| 44 |
+
## Honesty
|
| 45 |
+
|
| 46 |
+
| Claim | Label | What-NOT |
|
| 47 |
+
|---|---|---|
|
| 48 |
+
| Table built in this package | REPORTED | V=64 d=12, not 3290×128 |
|
| 49 |
+
| Neural / trained embed | FALSE | hash+table, not SGD |
|
| 50 |
+
| Analogy / retrieval score | UNAVAILABLE | not measured |
|
| 51 |
+
| Energy | UNAVAILABLE | never a fabricated joule |
|
| 52 |
+
| CUDA | UNAVAILABLE | CPU numpy LIVE |
|
| 53 |
+
|
| 54 |
+
Doctrine v11 LOCKED · 749/14/163 · locked-proven 8. Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID [0009-0001-0110-4173](https://orcid.org/0009-0001-0110-4173).
|