WangKaiLin/PipeOwl-1.5-jp

PipeOwl-1.5-Japanese (Geometric Embedding)

A transformer-free semantic retrieval engine.

PipeOwl performs deterministic vocabulary scoring over a static embedding field:

score = α⋅base + β⋅Δfield

where:

• base = cosine similarity in embedding space
• Δfield = static scalar field bias

Features:

• O(n) over vocabulary.
• No attention.
• No transformer weights.
• CPU-friendly (<40MB model)

Architecture

• Static embedding table (V × D)
• Aligned vocabulary index
• Optional scalar bias field (Δfield)
• Linear scoring
• Pluggable decoder stage
• Targeted for CPU environments and low-latency systems (e.g. IME).

Model Specs

item	value
vocab size	26155
embedding dim	768
storage format	safetensors (FP16)
model size	~38.7 MB
languages	Japanese
startup time	<1s
query latency	~3-4 ms (CPU, full vocabulary scan)

Attribution

DATA_SOURCES.md

Quickstart

git clone https://huggingface.co/WangKaiLin/PipeOwl-1.5-jp
cd PipeOwl-1.5-jp

pip install numpy safetensors

python quickstart.py

Example:

Example semantic retrieval results:

Please enter words： 日

Top-K Tokens:
0.974 | 日
0.794 | 日の
0.789 | 翌日
0.777 | 週
0.775 | 週間

Please enter words： 行

Top-K Tokens:
0.961 | 行
0.794 | 行こ
0.787 | 執り行
0.787 | 入
0.784 | 起

Please enter words： 東京

Top-K Tokens:
0.979 | 東京
0.872 | 大阪
0.868 | 名古屋
0.849 | 横浜
0.848 | 目黒

Benchmark (CPU)

Environment:

• Vocab size: 26,155
• Embedding dimension: 768
• Hardware: CPU

Average query latency:

• PipeOwl: 0.0036 sec
• BM25: 0.0421 sec
• Embedding: 0.0283 sec
• FAISS Flat: 0.0324 sec
• FAISS HNSW: 0.0230 sec

Comparison	Speedup
vs BM25	11.7× faster
vs Embedding	7.9× faster
vs FAISS Flat	9.0× faster
vs FAISS HNSW	6.4× faster

PipeOwl shows 6–12× lower latency compared with common retrieval baselines in this setup.

repo: https://huggingface.co/datasets/WangKaiLin/pipeowl-1.5-jp-benchmark

Repository Structure

pipeowl-1.5-jp/
 ├ README.md
 ├ config.json
 ├ DATA_SOURCES.md 
 ├ LICENSE
 ├ quickstart.py
 ├ engine.py
 ├ vocabulary.json
 └ pipeowl_fp16.safetensors

LICENSE

MIT