Update README.md

README.md

@@ -9,16 +9,18 @@ tags:
 
 # maid-yuzu-v8-alter-iMat-GGUF
 
-<b>Highly requested model.</b> Quantized from fp16 with love.
+<b>Update:</b> Legacy quants calculated with imatrix showed lower average divergence than expected when compared to their non-imat variants. Uploading those now as well.
 
-* <s>1st batch (IQ3_S, IQ3_XS) use a imatrix.dat file calculated from Q8 quant using an input file from [this discussion](https://github.com/ggerganov/llama.cpp/discussions/5263#discussioncomment-8395384)</s> These have been removed in favor of later method. Please see tables below.
+<b>Highly requested model.</b> Quantized from fp16 with love.
+* <s>1st batch (IQ3_S, IQ3_XS) use a imatrix.dat file calculated from Q8 quant.</s> These have been removed in favor of a newer method. Please see tables below.
 * Later files made using .imatrix file from [this](https://huggingface.co/datasets/ikawrakow/imatrix-from-wiki-train) repo (special thanks to [ikawrakow](https://huggingface.co/ikawrakow) again)
 
-<h3>Original IQ3_XS KL Divergence (calculated vs Q8)</h3>
+<h3>Original IQ3_XS KL-Divergence (calculated vs Q8)</h3>
 <img src="https://huggingface.co/InferenceIllusionist/maid-yuzu-v8-alter-iMat-GGUF/resolve/main/IQ3_XS_OLD_KL.JPG?download=true" width="300"/>
 
-<h3>Updated IQ3_XS KL Divergence (calculated vs Q8)</h3>
+<h3>Updated IQ3_XS KL-Divergence (calculated vs Q8)</h3>
 <img src="https://huggingface.co/InferenceIllusionist/maid-yuzu-v8-alter-iMat-GGUF/resolve/main/IQ3_XS_NEW_KL.JPG?download=true" width="300"/>
+Lower numbers are better
 
 For a brief rundown of iMatrix quant performance please see this [PR](https://github.com/ggerganov/llama.cpp/pull/5747)