← Dispatch

New imatrix Dataset for Better GGUF Quantization

2026-08-18 · Nightwing · 1 min read

What shipped? Bartowski released calibration_datav6.txt — a new importance matrix (imatrix) dataset for GGUF quantization, built from scratch with formatted text and chat templates.

What changed? The previous calibration_datav5.txt was plain text with no structure. The new dataset (v6) includes chat-templated conversations, tool-use examples, and formatted prose — all rendered through the model's own chat template before being fed into the imatrix calculation. The result is a calibration corpus that better reflects how models actually get used in production: structured conversations, not just raw text.

Bartowski tested dozens of alternative datasets against v5. The headline finding: v5 was already surprisingly good. But v6 edges ahead on structured tasks, particularly tool calling, where the formatted examples give the imatrix a more accurate signal for those weight groups.

Why does a builder care? If you're running local models via llama.cpp or any GGUF-based inference engine, the imatrix you use determines which weights get preserved during quantization. A better imatrix means better quality at the same model size — fewer surprises when your quantized Qwen3.8-27B tries to call a tool.