Higher than usual refusal rate with Q6_K_L quant GGUF
I've been running a number of different GLM 4 32b quants over the past few days, thought I'd try your new q6_k_l GGUF and interestingly - of all the GGUFs I've tried recently - it's the only one that outright refuses to create a SVG of a pelican riding a bicycle.
I am aware that a SVG of a pelican riding a bike is not exactly a 'measure of SOTA' π, but it is an interesting test - and it's really got me a bit stumped as to why this one in particular might be more stubborn or perhaps "safe".
Each GGUF I've used is run up in Ollama with the exact same inference parameters (temperature 0.2, top_p 1.0, min_p 0.05)
vs
That's so strange lmao.. I can't possibly begin to imagine why it would do that π
i'm running GLM-4-32B-0414-UD-Q3_K_XL by unsloth, but my settings are top_p 95, min_p 0, top_k 0, temp 0.6. First try got this result.
And overall i'm impressed with all the coding results so far.
Second attempt wasn't too good, but then 3rd one:
Promt is exactly as yours: Generate an SVG of a pelican riding a bicycle