GGUF quantizer bartowski released an improved calibration (‘imatrix’) dataset, version 6, used to guide which weights to preserve during LLM quantization by measuring how strongly each input channel activates on different text corpora. Testing across multiple models found that dataset choice barely affects quality above 4 bits per weight, but matters significantly at lower bitrates such as Q2_K, particularly for mixture-of-experts models that need full expert coverage. The final v6 dataset mixes diverse prose with chat-templated tool-calling data at a 2:3 ratio to balance broad coverage against measured performance.
