Update README.md

README.md

@@ -20,11 +20,10 @@ Original model: https://huggingface.co/google/gemma-2-27b-it
 
 All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
 
-Experimental quants are made with `--output-tensor-type f16 --token-embedding-type f16` per [ZeroWw](https://huggingface.co/ZeroWw)'s suggestion, please provide any feedback on quality differences you spot.
-
 ## What's new
 
 - June 21 2024: Contains latest tokenizer fixes, which addressed a few oddities from the original fix, should be closest to correct performance yet. Also has metadata for SWA and logit softcapping.
+- July 3 2024: Updated the experimental quants to newer method, Q8 for embed/output, yields higher quality at much lower size than f16 (left Q8_0_L since Q8_0 is already Q8 embed/output)
 
 ## Prompt format
 
@@ -43,27 +42,34 @@ Note that this model does not support a System prompt.
 | -------- | ---------- | --------- | ----------- |
 | [gemma-2-27b-it-Q8_0_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q8_1.gguf) | Q8_0_L | 30.04GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Extremely high quality, generally unneeded but max available quant. |
 | [gemma-2-27b-it-Q8_0.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q8_0.gguf) | Q8_0 | 28.93GB | Extremely high quality, generally unneeded but max available quant. |
-| [gemma-2-27b-it-Q6_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q6_K_L.gguf) | Q6_K_L | 23.73GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Very high quality, near perfect, *recommended*. |
+| [gemma-2-27b-it-Q6_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q6_K_L.gguf) | Q6_K_L | 22.62GB | Uses Q8_0 for embed and output weights. Very high quality, near perfect, *recommended*. |
 | [gemma-2-27b-it-Q6_K.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q6_K.gguf) | Q6_K | 22.34GB | Very high quality, near perfect, *recommended*. |
-| [gemma-2-27b-it-Q5_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q5_K_L.gguf) | Q5_K_L | 20.79GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. High quality, *recommended*. |
+| [gemma-2-27b-it-Q5_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q5_K_L.gguf) | Q5_K_L | 19.69GB | Uses Q8_0 for embed and output weights. High quality, *recommended*. |
 | [gemma-2-27b-it-Q5_K_M.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q5_K_M.gguf) | Q5_K_M | 19.40GB | High quality, *recommended*. |
 | [gemma-2-27b-it-Q5_K_S.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q5_K_S.gguf) | Q5_K_S | 18.88GB | High quality, *recommended*. |
-| [gemma-2-27b-it-Q4_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q4_K_L.gguf) | Q4_K_L | 18.03GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Good quality, uses about 4.83 bits per weight, *recommended*. |
+| [gemma-2-27b-it-Q4_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q4_K_L.gguf) | Q4_K_L | 16.93GB | Uses Q8_0 for embed and output weights. Good quality, uses about 4.83 bits per weight, *recommended*. |
 | [gemma-2-27b-it-Q4_K_M.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q4_K_M.gguf) | Q4_K_M | 16.64GB | Good quality, uses about 4.83 bits per weight, *recommended*. |
 | [gemma-2-27b-it-Q4_K_S.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q4_K_S.gguf) | Q4_K_S | 15.73GB | Slightly lower quality with more space savings, *recommended*. |
 | [gemma-2-27b-it-IQ4_XS.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ4_XS.gguf) | IQ4_XS | 14.81GB | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
-| [gemma-2-27b-it-Q3_K_XL.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q3_K_XL.gguf) | Q3_K_XL | 15.91GB | *Experimental*, uses f16 for embed and output weights. Please provide any feedback of differences. Lower quality but usable, good for low RAM availability. |
+| [gemma-2-27b-it-Q3_K_XL.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q3_K_XL.gguf) | Q3_K_XL | 14.80GB | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
 | [gemma-2-27b-it-Q3_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q3_K_L.gguf) | Q3_K_L | 14.51GB | Lower quality but usable, good for low RAM availability. |
 | [gemma-2-27b-it-Q3_K_M.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q3_K_M.gguf) | Q3_K_M | 13.42GB | Even lower quality. |
 | [gemma-2-27b-it-IQ3_M.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ3_M.gguf) | IQ3_M | 12.45GB | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
 | [gemma-2-27b-it-Q3_K_S.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q3_K_S.gguf) | Q3_K_S | 12.16GB | Low quality, not recommended. |
 | [gemma-2-27b-it-IQ3_XS.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ3_XS.gguf) | IQ3_XS | 11.55GB | Lower quality, new method with decent performance, slightly better than Q3_K_S. |
 | [gemma-2-27b-it-IQ3_XXS.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ3_XXS.gguf) | IQ3_XXS | 10.75GB | Lower quality, new method with decent performance, comparable to Q3 quants. |
+| [gemma-2-27b-it-Q2_K_L.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q2_K_L.gguf) | Q2_K_L | 10.73GB | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
 | [gemma-2-27b-it-Q2_K.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-Q2_K.gguf) | Q2_K | 10.44GB | Very low quality but surprisingly usable. |
 | [gemma-2-27b-it-IQ2_M.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ2_M.gguf) | IQ2_M | 9.39GB | Very low quality, uses SOTA techniques to also be surprisingly usable. |
 | [gemma-2-27b-it-IQ2_S.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ2_S.gguf) | IQ2_S | 8.65GB | Very low quality, uses SOTA techniques to be usable. |
 | [gemma-2-27b-it-IQ2_XS.gguf](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF/blob/main/gemma-2-27b-it-IQ2_XS.gguf) | IQ2_XS | 8.39GB | Very low quality, uses SOTA techniques to be usable. |
 
+## Credits
+
+Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset
+
+Thank you ZeroWw for the inspiration to experiment with embed/output
+
 ## Downloading using huggingface-cli
 
 First, make sure you have hugginface-cli installed: