Sampling methods also affect this. Have you tried min_p sampling? https://github.com/ggerganov/llama.cpp/pull/3841#issuecommen...