Perplexity windows versus a task check for the 2-bit Qwen3.5-2B files
#1
by guicybercode - opened
Publishing per-window NLL so two files can be compared pairwise is more useful than a single perplexity number. For Qwen3.5-2B, unsloth UD-IQ2_XXS lands at 23.41 against 13.33 for DASH-Q IQ2_M at similar file sizes. Did you run any task check on the 2B files, even a small one? A recent paper (Broken on Arrival, 2609.05881) found low-bit tags of models up to 4B that were non-functional on code while the files themselves were intact, and I would like to know if the 2B file with the worst perplexity here also fails on tasks or only drifts. Also, were any of the windows run fully on CPU, or only the rival tensors placed there with -ot?