yaoandy107/greyscope-v2-qwen3.5-4b-int4

Greyscope v2 int4

This is the 3.5 GB Transformers int4-HQQ build of Greyscope v2. It returns the same continuous ai_involvement score and human / AI-edited / AI-generated labels as the bf16 model.

Use this build when you need Transformers but cannot fit bf16. Do not use it on Apple Silicon; use MLX 4-bit.

Quick start

git clone https://github.com/yaoandy107/greyscope
cd greyscope
uv sync --extra int4
uv run greyscope --model int4 "Paste a paragraph here."

Quantization check

The release export checked this artifact against bf16 on 64 rows. Treat this as a smoke test, not a quality benchmark, and validate it on your own data and hardware.

Evaluations, calibration, training details, and limitations are documented on the bf16 model card.