hellaswag: by models

Home Paper Code


SE predicted by accuracy

The typical standard errors between pairs of models on this dataset as a function of the absolute accuracy.

SE vs accuracy difference

The standard error of each model pair against their observed accuracy difference. Pairs below a reference line differ by more than the corresponding number of standard errors, i.e. they are statistically distinguishable at that level.

CDF of question level accuracy

Results table by model

model pass1 pass@count win_rate count SE(A) SE_x(A) SE_pred(A)
dbrx-base 88.7 88.7 13.9 1 0.32 NaN NaN
Mixtral-8x22B-v0.1 86.8 86.8 11.7 1 0.34 NaN NaN
Qwen1.5-110B 86.5 86.5 11.5 1 0.34 NaN NaN
Meta-Llama-3-70B 85.9 85.9 11.1 1 0.35 NaN NaN
deepseek-llm-67b-base 85.5 85.5 10.5 1 0.35 NaN NaN
Qwen1.5-72B 85.3 85.3 10.8 1 0.35 NaN NaN
llama_65B 85.3 85.3 10.4 1 0.35 NaN NaN
falcon-40b 85.1 85.1 10.4 1 0.36 NaN NaN
Mixtral-8x7B-v0.1 84.5 84.5 10 1 0.36 NaN NaN
Qwen1.5-32B 84.1 84.1 9.85 1 0.36 NaN NaN
llama_33B 84 84 9.63 1 0.37 NaN NaN
llama2_70B 83 83 9.82 1 0.37 NaN NaN
Mistral-7B-v0.1 81.7 81.7 8.22 1 0.39 NaN NaN
gemma-7b 80.8 80.8 7.84 1 0.39 NaN NaN
mpt-30b 80.8 80.8 7.77 1 0.39 NaN NaN
Meta-Llama-3-8B 80.5 80.5 7.47 1 0.4 NaN NaN
llama_13B 80.4 80.4 7.5 1 0.4 NaN NaN
llama2_13B 80.3 80.3 8.39 1 0.4 NaN NaN
Qwen1.5-14B 80 80 7.71 1 0.4 NaN NaN
deepseek-moe-16b-base 78.6 78.6 6.65 1 0.41 NaN NaN
falcon-7b 78.3 78.3 6.58 1 0.41 NaN NaN
Qwen1.5-7B 77.3 77.3 6.47 1 0.42 NaN NaN
deepseek-llm-7b-base 77.2 77.2 6.01 1 0.42 NaN NaN
llama_07B 77.1 77.1 6.02 1 0.42 NaN NaN
llama2_07B 76.2 76.2 6.64 1 0.42 NaN NaN
stablelm-base-alpha-7b-v2 75.5 75.5 5.35 1 0.43 NaN NaN
stablelm-3b-4e1t 75.2 75.2 5.24 1 0.43 NaN NaN
gemma-2b 71.7 71.7 4.24 1 0.45 NaN NaN
Qwen1.5-4B 71.6 71.6 4.46 1 0.45 NaN NaN
pythia-12b-deduped-v0 69.5 69.5 3.71 1 0.46 NaN NaN
pythia-6.9b-deduped-v0 66.1 66.1 2.96 1 0.47 NaN NaN
Qwen1.5-1.8B 61 61 2.2 1 0.49 NaN NaN
pythia-2.8b-deduped 60.3 60.3 1.97 1 0.49 NaN NaN
pythia-1.4b-deduped-v0 52 52 1.35 1 0.5 NaN NaN
pythia-1b-deduped 49.6 49.6 0.949 1 0.5 NaN NaN
Qwen1.5-0.5B 49.4 49.4 1.07 1 0.5 NaN NaN