The exact 512-sample calibration blend used to GPTQ-quantize canada-quant/hy3-w4a16-mtp (a W4A16 quantization of tencent/Hy3). Published for full reproducibility of the quantization pipeline.

Why a blend (not chat-only)

INT4 weight quantization degrades most on code and tool-call-shaped tokens. A chat-only calibration set (e.g. pure ultrachat) under-samples exactly the routed experts those tokens activate. This set deliberately… See the full description on the dataset page: https://huggingface.co/datasets/canada-quant/hy3-w4a16-mtp-calibration.