gsm8k_cot: 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)
google_gemma_3_27b_it 94.4 95.9 32.6 3 0.63 0.56 0.29
qwen3-14b 93.5 95.6 32.2 3 0.68 0.59 0.34
qwen3-32b 93.4 95.8 32.2 2 0.69 0.53 0.43
llama-3.1-70B-instruct 92.9 92.9 31.9 4 0.71 0.71 0
google_gemma_3_12b_it 92.3 95.6 31.4 4 0.73 0.61 0.4
qwen3-8b 92 94.8 31 3 0.75 0.63 0.4
qwen2.5-coder-32b-instruct 91.3 95.1 30.8 2 0.78 0.56 0.54
deepseek_r1_distill_llama_70b 90.6 94.8 30.9 2 0.8 0.57 0.56
google_gemma_2_27b_it 89.8 92.8 29.6 2 0.83 0.68 0.48
qwen2-72b-instruct 89.5 94.8 30.1 2 0.84 0.56 0.63
qwen3-4b 89.5 92.9 29.6 4 0.85 0.76 0.38
google_gemma_2_9b_it 86.8 92.8 28 3 0.93 0.74 0.57
qwen2-math-72b-instruct 84.4 93.9 27.7 2 1 0.52 0.85
google_gemma_3_4b_it 83.3 91.1 26.3 5 1 0.86 0.57
qwen1.5-72b-chat 82.7 90.1 26.4 2 1 0.72 0.75
deepseek_r1_distill_qwen_7b 81.9 94.5 26.3 4 1.1 0.67 0.82
qwen1.5-32b-chat 81.7 90.1 26 2 1.1 0.7 0.8
deepseek_r1_distill_qwen_14b 81.3 93.3 26.4 4 1.1 0.72 0.8
qwen2.5-coder-14b-instruct 81 95.6 26.1 3 1.1 0.58 0.91
deepseek_r1_distill_qwen_32b 80.3 91.3 26.8 2 1.1 0.61 0.91
qwen2-math-7b-instruct 79.8 93.3 24.8 4 1.1 0.74 0.82
llama-3.1-8B-instruct 78.3 78.3 23.8 7 1.1 1.1 0
mistralai_mixtral_8x22b_instruct_v0.1 78.3 89.3 24.4 2 1.1 0.67 0.91
qwen2-math-1.5b-instruct 75.2 90.6 22.7 4 1.2 0.81 0.87
mistralai_ministral_8b_instruct_2410 73.9 92.3 22.3 4 1.2 0.74 0.96
qwen3-1.7b 73.4 84.8 21.6 4 1.2 0.99 0.71
deepseek_r1_distill_llama_8b 73 92.9 22.5 4 1.2 0.71 1
qwen2-7b-instruct 71.7 91.9 21.6 4 1.2 0.73 1
qwen1.5-14b-chat 70.4 85.5 20.3 3 1.3 0.87 0.9
llama-3.2-3B-instruct 67.6 67.6 19.4 10 1.3 1.3 0
qwen2.5-coder-7b-instruct 65.4 89.5 19.2 4 1.3 0.74 1.1
mistralai_mathstral_7b_v0.1 63.4 89.8 18.2 4 1.3 0.73 1.1
deepseek_v2_lite_chat 62.1 75.2 17.1 2 1.3 0.89 1
mistralai_mixtral_8x7b_instruct_v0.1 60.8 81.9 17.1 3 1.3 0.84 1
deepseek_r1_distill_qwen_1.5b 59.8 86.6 17.3 4 1.4 0.72 1.1
qwen1.5-7b-chat 55.2 75.4 14.5 3 1.4 0.91 1
qwen2.5-coder-3b-instruct 54.9 82.1 14.6 4 1.4 0.81 1.1
google_codegemma_1.1_7b_it 50.3 79 13 5 1.4 0.89 1.1
google_gemma_3_1b_it 44.6 62.9 10.9 4 1.4 1.1 0.86
mistralai_mistral_7b_instruct_v0.3 43.4 72.2 11 4 1.4 0.82 1.1
qwen3-0.6b 38.2 64.8 8.87 5 1.3 0.92 0.98
qwen2.5-coder-1.5b-instruct 37.4 64.8 8.72 4 1.3 0.82 1
mistralai_mistral_7b_instruct_v0.2 36.8 62.4 8.82 4 1.3 0.86 1
llama-3.2-1B-instruct 29 29 6.29 13 1.2 1.2 0
google_gemma_7b_it 27.5 47.5 6.17 4 1.2 0.85 0.89
mistralai_mistral_7b_instruct_v0.1 22.5 47.8 4.79 4 1.1 0.63 0.96
qwen2-1.5b-instruct 20.5 47.8 4.43 4 1.1 0.53 0.98
qwen1.5-1.8b-chat 11.1 24.5 2.12 3 0.86 0.41 0.76
qwen2.5-coder-0.5b-instruct 9.83 29.8 1.87 5 0.82 0.38 0.73
google_gemma_2b_it 9.17 21.1 1.77 4 0.79 0.49 0.62
qwen2-0.5b-instruct 8.49 27.1 1.58 5 0.77 0.33 0.69
qwen1.5-0.5b-chat 3.12 12.7 0.622 5 0.48 0.14 0.46