ARotting's picture
Publish Reproducible low-bit classification split manifest
e4d414b verified
|
Raw
History Blame Contribute Delete
1.58 kB
metadata
license: apache-2.0
tags:
  - quantization
  - binary-neural-network
  - ternary-weights
  - knowledge-distillation
  - gradio

BitForge 1-bit

BitForge trains and compares a full-precision digit classifier, a strict one-bit weight model, and a ternary-weight model. The low-bit students use straight-through quantization during training and learn from both labels and the full-precision teacher's softened output distribution.

The binary deployment artifact stores each matrix weight as one packed sign bit, plus one floating-point scale per output channel and floating-point biases. The project reports both classification accuracy and the measured inference payload. Activations, scales, and biases remain floating point, so this is specifically a one-bit matrix-weight experiment rather than a claim that every operation or parameter is one bit.

Verified results

Variant Test accuracy Accuracy change
FP32 teacher 95.78% reference
Packed binary matrix weights 94.22% -1.56 points
Ternary matrix weights 95.11% -0.67 points

Each network has 4,810 parameters, including 4,736 matrix weights. The measured inference payload fell from 19,240 bytes for FP32 parameters to 1,184 bytes for packed signs, per-channel scales, and biases, a 16.25 times reduction. The .npz container itself is 3,270 bytes because it also carries names, shapes, and archive metadata. An independent reload of the packed signs reproduced 94.22% accuracy.

Reproduce

uv run python projects/bitforge-1bit/train.py