Limour/perplexity

https://www.kaggle.com/code/reginliu/perplexity

Model	Size	PPL	n_vocab	PPL_adjust
qwen2.5-14b-fp16.gguf	27.51	9.5316 +/- 0.08886	152064	9.5316
qwen2.5-14b-IQ4_XS.gguf	7.56	9.6508 +/- 0.09039	152064	9.6508
qwen1_5-14b-chat-IQ3_XS.gguf	6.48	11.8084 +/- 0.121615	152064	11.8084
causallm_14b.IQ3_XS.gguf	6.48	13.3798 +/- 0.13641	152064	13.3798
causallm_14b.IQ4_XS.gguf	7.85	13.4127 +/- 0.13762	152064	13.4127
causallm_14b.Q4_0.gguf	8.18	13.6714 +/- 0.13964	152064	13.6714
causallm_14b.IQ2_XXS.gguf	4.98	15.0160 +/- 0.15004	152064	15.0160
Yi-9B-200K_iQ3xxs.gguf	3.47	6.8157 +/- 0.05453	64000	16.1941
Fi-9B-200K-Q8_0.gguf	9.38	6.8402 +/- 0.05741	64000	16.2523
causallm_7b.Q5_K_M.gguf	5.53	16.5278 +/- 0.18005	152064	16.5278
Qwen1.5-22B-Chat-Merge-Q4_0.gguf	12.6	21.9669 +/- 0.28980	152064	21.9669
Kunoichi-DPO-v2-7B-Q4_K_M-imatrix.gguf	4.37	6.7096 +/- 0.04519	32000	31.8840
WizardLM-2-7B-IQ4_XS-imat.gguf	3.91	9.8891 +/- 0.08106	32000	46.9930

For a model that returns tokens completely at random, we have $$P(token\mathrm{\mid }context)=\frac{1}{n_{vocab}},PPL=\sqrt[N]{{\left(\frac{1}{P}\right)}^N}=n_{vocab}P(token|context)\; =\; \backslash frac\{1\}\{n\_\{vocab\}\},\; \backslash quad\; PPL\; =\; \backslash sqrt[N]\{\backslash left(\backslash frac\{1\}\{P\}\backslash right)^N\}\; =\; n\_\{vocab\}$$ therefore $$PP{L}_{adjust}=\frac{PPL}{n_{vocab}}\times 152064PPL\_\{adjust\}\; =\; \backslash frac\{PPL\}\{n\_\{vocab\}\}\; \backslash times\; 152064$$