Downloads · 30 days
0
lossless-ternary
operajdw/lossless-ternary
lossless-ternary is a machine learning model from operajdw. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Lossless Base-3/5 + Gretchen — a multiply-free ternary toolkit
Access
Public
Updated Jul 10, 2026
Repo size
—
Likes
1
Public
Repo makeup
Click a slice to open those files.
Other1.5 KB · 58%
At a glance
- License
- mit
- Access
- Public
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026
- SHA
- 669af1e7
Model card
From the Hugging Face model README
Lossless Base-3/5 + Gretchen — a multiply-free ternary toolkit
Two pieces that fit together:
- Gretchen — train LoRA weights that crystallize to ternary {-1, 0, +1} during training, via a smooth polynomial well. No rounding, no straight-through estimator.
- Lossless base-3/5 kernels — represent values exactly in base-3/5 (a change of base, zero rounding) and run ternary matmuls multiply-free: add / skip / subtract, no floating-point multiply, no GPU, no CUDA. Proven bit-exact in Python (torch.allclose, atol=0) and in framework-free C.
The case: a ternary model's matmul is compute-light and memory-light, so it belongs on a CPU with SIMD or hand-written assembly, not a GPU built for floating-point multiplies. (Cf. Microsoft's bitnet.cpp.) In the spirit of ancient Egyptian arithmetic — multiply by shift-and-add, represent exactly by addition, never round.
Code, tests, and full write-ups: https://github.com/drogongod/lossless-ternary
Author: Jonathan David Wint. License: MIT.
- License
- mit
- Languages
- en
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026