Open Hermes + Starling passthrough merged
SlimOrca(?)+Zephyr Beta linear merged, then passthrough merged with Synthia
Then both models were merged again in 1 to 0.3 ratio.
My findings:
Increasing repetition penalty usually makes the model smarter up to a degree but it also causes stability issues.
Since most of the merged models were trained with ChatML, use ChatML template. Rarely the model throws another EOS token though.
- My favorite preset has been uploaded.
- You can use some sort of CoT prompt instead of "system" in ChatML. It does improve the quality of most output.
(You are an assistant. Break down the question and come to a conclusion.)
I don't know what I am doing, you are very welcome to put the model through benchmarks.
I'll also upload q6 GGUF but my internet is shit, so don't hesitate to share other quantizations.
