Yaz — an editable, auditable tiny knowledge model that abstains when unsure

Yaz is a sub-1M-parameter (≈808K), byte-level language model whose individual facts you can create, read, update, and delete one at a time — with provable per-edit locality — and that abstains when it isn't confident which fact you mean, instead of guessing. CPU-only, offline.

Status: research prototype. Small-scale and honestly scoped. A clean, reproducible demonstration — not a state-of-the-art result and not a defensible new capability. Read the limitations below.

📄 Technical report: paper/yaz-technical-report.pdf
💻 Code & reproduction: https://github.com/TilelliLab/Yaz
🔬 Each fact = one decoder column (atom); routing by a frozen all-MiniLM-L6-v2 embedding.

How it works

Each fact lives in its own addressable atom (a decoder column). A frozen sentence embedding routes a prompt to a fact by meaning, so paraphrases reach the same fact. UPDATE swaps a column, DELETE zeroes it, CREATE allocates a fresh one, READ is just routing — no retraining. The routing confidence margin (top-1 − top-2) is used as an "I don't know which fact you mean" signal, so the model refuses low-confidence queries.

Download

Use the Hugging Face Hub (handles the Git-LFS weights for you). Don't git clone without git-lfs — you'd get 132-byte LFS pointer files instead of the real model.safetensors / .pt.

bash

pip install huggingface_hub
# CPU-only deps (avoids the multi-GB CUDA stack):
pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu

python

from huggingface_hub import snapshot_download
repo = snapshot_download("TilelliLab/Yaz")     # -> local path with all files (real weights)
# then: cd into `repo` and run the snippets below, or `python demo.py --demo`.

Load it (safetensors, no pickle)

python

# files in this repo: model.safetensors, yaz_meta.json, load_yaz.py, yaz/ (model code)
from load_yaz import load_yaz
model, cfg, meta = load_yaz()          # 807,680 params, 50 fact-atoms
print(meta["country_to_target_atom"]["France"])   # -> 0

Run the full routing + abstention + live-edit demo (needs pip install -r requirements.txt):

bash

python demo.py --demo
python demo.py --prompt "the country of the Eiffel Tower, its capital is "
python demo.py --prompt "The capital of France is " --edit France=Lima
python demo.py --prompt "best pizza topping?"        # -> ABSTAIN (out of scope)

The original PyTorch checkpoint is also included at checkpoints/yaz_gen_semantic_v2.pt for fidelity; the model.safetensors is the recommended (pickle-free) artifact.

What it can do (measured)

Capability	Result
UPDATE (edit, no retraining)	in-dist reliability 1.000; edits land 8/8 (first byte)
DELETE	fact gone, 0 collateral
CREATE	passes 4/4 battery (monosemantic / local / readable / deletable)
Per-edit locality	0/10 collateral; bpc +0.000% across 40 sequential edits
Paraphrase-robust routing	held-out reach 0.696 vs 0.216 surface routing
Abstain when unsure	near-oracle risk-coverage AURC 0.004 (oracle 0.003)

All numbers reproduce with the public all-MiniLM-L6-v2 embedder (no internal dependencies), seed 2026, CPU.

How it compares

Yaz sits in the side-memory family of editors (like SERAC/GRACE/WISE/PENME), not the weight-editing family (ROME/MEMIT/MEND). Two honest differences: its edit is a structural decoder-column swap (locality by construction), and it abstains on low routing confidence instead of answering with the base model — a feature no published editor has. On a controlled tiny task, weight-editing (ROME/MEMIT) lands the new answer but loses selectivity as edits accumulate (22–67% of other facts change), while Yaz holds 0%; but Yaz is far smaller-scale and its update reliability degrades past its capacity. Full, honest comparison (incl. ROME/MEMIT-vs-Yaz at 5/50/200 facts): COMPARISON.md on GitHub.

Limitations (read these)

First-byte editor. Edits set the answer's first byte; multi-byte generation is not faithful (full-word transfer ≈ 0.05).
A retracted claim. An earlier "edit-generalization" headline of 0.675 was retracted — a random-column-swap control sits at ≈ 0.688, i.e. that number was at chance. What survives is routing reach, not an edit-magnitude effect.
Fragile routing on oblique, name-free clues (≈0.85 famous → ≈0.50 oblique).
Structural locality holds only while no two facts share an atom.
Tiny, synthetic scope — 50 country→capital facts, single seed, CPU.
Not a moat. Mechanisms exist in the literature (ROME/MEMIT, GRACE, SERAC, PENME; EasyEdit). Yaz combines them cleanly and reproducibly — an engineering contribution, not a unique capability.

TilelliLab/
Yaz