Hope you are okay
I feel like I am the HF version of The Ring.. I follow you and in 7 days you disappear. Hope you are okay.
Hey, thank you for your concern. I am ok. I have just been very busy writing music and trying to launch my music career. I have not had much time for anything else. So far, it is going well, as I have been noticed and contacted by other artists, and I have now work in progress to write and compose songs for 3 big groups/artists.
On the LLM side, I am still using them regularly. I have not done any official benchmark, but I still have not found anything local that can beat WizardLM 8x22b. My latest quick tests were with DeepSeek R1 Distilled.
DeepSeek-R1-Distilled-Qwen-32b-bf16: loses the plot very quickly and starts repeating itself. Bad writing quality. Typical of the Qwen serie.
DeepSeek-R1-Distilled-Llama-70b-8bit: horrible writing quality; not great at following instructions and reasoning. Typical of the Llama serie, which is not surprising when the source data is such a heap of rubbish.
I feel that there has been so much progress in the recent month with online offerings, I tend to use those more than before. I still rely on local LLM from time to time, but less so. Mostly when I need something uncensored and objective.
Great to hear you are doing well! I wish you much success and joy from your musical pursuits! 🙏
I still have not found anything local that can beat WizardLM 8x22b
I know you probably can't run it locally but the full R1 might top your charts if you end up benchmarking it. The distill's aren't very representative of the true R1. Deepseek V3 is in my opinion not worth using, R1 is much better and it also has another benefit over V3 because having and being able to mold the thinking tokens enhances steerability. I'm running it locally with a 354 GiB quant.
Edit: I noticed you test with repeat_penalty, I think this hurts model performance, and the problem it exists to solve either doesn't exist in modern LLM or can be dealt with better such as with DRY.