Update README.md

README.md

@@ -16,6 +16,29 @@ datasets:
 - Locutusque/function-calling-chatml
 - internlm/Agent-FLAN
 ---
+# GGUF fix
+This is the same model as [cognitivecomputations/dolphin-2.9.1-dbrx](https://huggingface.co/cognitivecomputations/dolphin-2.9.1-dbrx), but with [gguf fixes made by Kenjoyer](https://huggingface.co/Kenjoyer/dolphin-2.9.1-dbrx-llamacppfix) applied(thanks a lot!). This model can be converted into gguf using llama.cpp.
+
+# Benchmarks and personal opinion
+### NeoEvalPlusN_benchmark
+[My meme benchmark.](https://huggingface.co/datasets/ChuckMcSneed/NeoEvalPlusN_benchmark)
+|Name                                         |Quant|Size  |B  |C  |D  |S   |P   |total|BCD|SP  |
+|---------------------------------------------|-----|------|---|---|---|----|----|-----|---|----|
+|cognitivecomputations/dolphin-2.9.1-dbrx     |Q6_K |16x12B|3  |1  |3  |4   |6   |17   |7  |10  |
+|cognitivecomputations/dolphin-2.9.1-qwen-110b|Q6_K |110B  |0  |1  |3  |3.75|4.25|12   |4  |8   |
+|databricks/dbrx-instruct                     |Q6_K |16x12B|0  |0  |0  |6.5 |4.5 |11   |0  |11  |
+|cognitivecomputations/dolphin-2.2-70b        |Q6_K |70B   |0  |1  |1  |4.5 |4.5 |11   |2  |9   |
+|Maximum        |n/a |n/a   |3  |2  |3  |8 |6 |22   |8  |14   |
+More compliant than the official instruct tune(BCD). To my surprise, performed much better overall than qwen-110b tuned on the same dataset. 
+Wrote 6 perfect poems(P column), which is **very** unusual. 
+Only models from goliath family and more recent llama-3-70b-instruct could do that. 
+Stylized writing tests(S column) were a bit disappointing, Dolphin is not famous for that. 
+In practical use, did perform better than the official tune. Still knows a lot, just like the official tune. 
+Writing is not great, wouldn't use it over Command-r+, unless I need to know some obscure facts. Feels like quantization hurts it a lot more than dense models.
+
+Verdict: Meh, just like the other dolphins. Eric, no disrespect, but you need to get better datasets. GPTslop really hurts practical performance.
+
+# Original model card below
 
 # Dolphin 2.9.1 DBRX 🐬