On wikitext2 there was no difference. I concluded these have no long-term dependency and I should use Linux kernel. Still, the same.
So yes, KLD depends on the dataset. Still, it does not measure what any e2e test does.
As for KLD, have you tried it on something which is even closer to e2e task, like agentic traces from https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-SF... or similar datasets?
If you would like to do so, it is easy (and orders of magnitude cheaper) than running benchmarks.
Thank you for calling this out. Using Wikipedia snippets for these is a terrible choice. I did a bunch of KL and other stats with the five Gemma 4 models, and the results were non-obvious. Anthropomorphizing:
Gemma 4 31B: "I guess we'll pretend I said this, but it's not me." (Baseline for stats)
Gemma 4 26B: "Dude, I'm certain I wouldn't have said this." (Bad KL)
Gemma 4 12B: "Umm, Me either!" (Similarly Bad KL)
Gemma 4 E4B: "I might say almost anything, this is fine." (Much better KL!!!)
Gemma 4 E2B: "I'm basically a toy. Let's play a game!" (Same KL as E4B)
Anyway, for comparing quantizations, it seems like the largest precision version should be given a one-shot prompt, and the result from that should be used as the corpus for the quantized versions.
Way better than wikitext-- but tells you nothing about errors tending to compound or cancel out.
Like say a test shows that only one token in a 10,000 token test would be different. Sounds very close, ship it!-- but what if trajectories with that single different token guarantees failure because it sets in motion a cascade of differences that ultimately result in a final distribution that doesn't include the solution?