Update README.md

README.md

@@ -7,8 +7,74 @@ library_name: transformers
 tags:
 - mergekit
 - merge
-license: llama3.2
+license: llama3.3
 ---
+# about
+
+The Teaz series is my third attempt at making merges, after the Kostume and Kermes series.
+
+This time, the goal was to make a smart model with a low perplexity, in accordance to the principles of the Kermes series, but with a merge of 3 merged models like on the kostume series.
+
+Huihui's abliterated models were used:
+- Llama 3.3 70b as the pivot of the first/main model.
+- Nemotron 3.1 70b and Deepseek R1 Distill 70b as the pillars.
+- and Tulu 3 70b as the backers of the 2nd and 3rd models.
+
+Bingo again. I hit 3.45 ppl512 wikieng, 62+ or ARC-C, and 82+ on ARC-E. Absolute top of the class for L3.x 70b, like Kermes is for L3 3.2 3b.
+
+No cheating, no contaminating, just the wonderful MergeKit model-stock merge technique leveraged to a new level (compared to what I already saw being done, at least).
+
+Next projects will involve that model as the "smarts pillar" of further merges, aimed at any use case.
+
+---
+# credits
+
+Kudos go to the model authors, and to the Arcee / MergeKit folks, as well as to HF hosting the MergeKit App.
+Also a big-up to SteelSkull, observing him cooking Nevoria decided me to try to make some merges by myself.
+
+---
+# historic
+
+1) On the Kostume series started on the 11/02/0205 I tried to make a triple stock merge of 3 intermediary stock merges of a dozen of model or so.
+This, to see if I could pile up their abilities.
+
+Not bad, but nothing special about it, it's a bit hard for me to judge at 3b.
+
+2) On the Kermes series started the day after, I defined a simpler approach:
+
+- Perplexity is the main constraint. Usual L3.2 3b finetunes are around 10.5-11 ppl512wikieng, Hermes is around 9.5.
+- I also measure in French and Serbian to observe the variances.
+  
+- Arc Challenge and Easy are the second constraint to judge its basic logics.
+- Usual L3.2 3b finetunes hit 40 and 60-65 respectively, Hermes3 hits 47+ and 70+.
+
+- Lack of censorship. I always keep in mind to pick models compatible with that AMAP.
+- This, may it be through the picked models' abliteration or the datasets they use.
+
+- And of course, the test, both In Kobold/Croco.CPP (spamming very offensive requests), and in ST (a 10k prompt with a big lorebook).
+
+Kermes series are basically stock merges on the top of anothers.
+- The goal was to maintain as much the qualities of the models used, so I stay on 1+2 models for the first merge, and 1+2 for the second as well.
+
+And bingo. Perplexity goes down still, ARC remain stable, it's quite unhinged still, and.. quite coherent, event at 10k+ context.
+
+---
+# quantizations
+
+GGUF static quantizations (Thanks Mradermacher!) :
+
+https://huggingface.co/mradermacher/Llama_3.2_3b_Kermes_v2.1-GGUF
+
+GGUF iMatrix quantizations (Thanks Mradermacher!) :
+
+https://huggingface.co/mradermacher/Llama_3.2_3b_Kermes_v2.1-i1-GGUF
+
+GGUF custom iMatrix quantizations: 
+
+https://huggingface.co/Nexesenex/Llama_3.2_3b_Kermes_v2.1-iMat-CQ-GGUF
+
+---
+# merge
 # merge
 
 This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).