DBF/SIGN1

Reproducible conversion and evaluation pipeline for post-training structural binarization of routed LLM experts using Double Binary Factorization (DBF) and the SIGN1 llama.cpp tensor type.

Status: private release-staging repository. Scripts and manifests are being frozen against the final model and public eodus/sign1-llama.cpp commit before publication.

Current result

Qwen3.6-35B-A3B routed experts, k=1024:

• DBF payload: 2.5546875 bits per original expert weight (slightly below Q2_K's 2.625 bpe).
• Full WikiText-2 PPL: FP16 experts 6.9935; Q2_K 7.7823; DBF 7.3186.
• DBF removes 57.5% of Q2_K's FP16 gap on the mean-NLL scale.
• Matched Q3_K PPL: 7.2279; DBF recovers 83.1% of the Q2_K-to-Q3_K improvement.
• Local scalar-equivalent interpolation: approximately 3.30 bpe quality at a 2.555 bpe DBF payload.
• Mean matrix residual improvement over Q2_K: 19.10% Frobenius and 22.52% spectral; DBF wins all 30,720 matrices in both metrics.
• Radeon 8060S Vulkan: DBF is 1.63% behind Q2_K in pp512 and 0.46% behind in tg128.

Repository layout

• scripts/run_dbf_all_experts.py — resumable 80-iteration fixed-dimension DBF core plus final middle-diagonal reprojection.
• scripts/analyze_dbf_all_experts.py — exact ggml Q2_K controls and whole-model Frobenius/spectral statistics.
• scripts/build_dbf_all_experts_model.py — final SIGN1 DBDBD GGUF builder and manifest.
• scripts/build_matched_scalar_experts.py — matched scalar quality controls.
• scripts/run_full_ppl.sh — paired full-PPL run with explicit path arguments.
• scripts/bench_abba.sh / scripts/parse_abba.py — drift-balanced final speed comparison.
• scripts/bratislava_admm.py — executable fixed-dimension DBF core adapted from the public Boža–Macko notebook.

Required companion runtime

The GGUF uses custom SIGN1 tensors and a routed D_in -> V -> D_mid -> U -> D_out graph. Stock llama.cpp cannot load it. Build the exact release commit from:

https://github.com/eodus/sign1-llama.cpp

Attribution

Double Binary Factorization was introduced by Vladimír Boža and Vladimír Macko:

• Paper: https://openreview.net/forum?id=k5kUKoewdQ
• Public reference code: https://github.com/usamec/double_binary

The SIGN1/DBDBD inference implementation, whole-model conversion, controls, and measurements were developed by Sasha Shlemov with Drinkins, his personal AI research and engineering assistant.

Publication checklist

• Freeze exact llama.cpp commit.
• Freeze model SHA-256 and GGUF manifest.
• Add Hugging Face model card and stage the repository privately before publication.
• Remove machine-specific public launchers; generic launchers take explicit paths.
• Add one top-level reproduction procedure (REPRODUCE.md).
• Verify entry points, shell syntax, and a three-matrix decomposition/metrics smoke from a fresh public checkout.
• Add Apache-2.0 license and upstream-code notices.