The typical stddev between pairs of models on this dataset as a function of the absolute accuracy.
Here is a more informative figure of the source information used to compute p-value. Any model pair to the right of the parabola is statistically different from each other at the given level. This plot shows a pretty sharp transition since there are no model pairs with a small #A_win + #B_win, which rules out significant results at a small difference in |#A_win-#B_win|. For more explanation see doc.
The null hypothesis is that model A and B each have a 1/2 chance to win whenever they are different, ties are ignored. The p-value is the chance under the null-hypothesis to get a difference as extreme as the one observed. For all pairs of models, the significance level mainly depends on the accuracy difference as shown here. Hover over each model pair for detailed information.
We show 3 methods currently used for evaluating code models, raw accuracy used by benchmarks, average win-rate over all other models (used by BigCode), and Elo (Bradly-Terry coefficients following Chatbot Arena). Average win-rate always have good correlation with Elo. GPT-3.5 gets an ELO of 1000 when available, otherwise the average is 1000. std: standard deviation due to drawing examples from a population, this is the dominant term. std_i: the standard deviation due to drawing samples from the model on each example. std_total: the total standard deviation, satisfying std_total^2 = std^2 + std_i^2.
| model | pass1 | std(E(A)) | E(std(A)) | std(A) | N | win_rate | elo |
|---|---|---|---|---|---|---|---|
| gpt-4-0613+cot | 77.2 | 1.2 | 0.81 | 1.5 | 10 | 44.2 | 1.12e+03 |
| gpt-4-0613 | 68 | 1.6 | 0.5 | 1.6 | 10 | 35.1 | 1.08e+03 |
| gpt-3.5-turbo-0613+cot | 56.5 | 1.4 | 1.1 | 1.8 | 10 | 24.7 | 1.03e+03 |
| deepseek-instruct-33b | 48.4 | 1.6 | 0.84 | 1.8 | 10 | 17.1 | 1e+03 |
| gpt-3.5-turbo-0613 | 47.5 | 1.6 | 0.78 | 1.8 | 10 | 17.1 | 1e+03 |
| deepseek-base-33b | 45.5 | 1.5 | 0.91 | 1.8 | 10 | 14.3 | 991 |
| codetulu-2-34b | 43.8 | 1.5 | 0.87 | 1.8 | 10 | 12.5 | 981 |
| codellama-34b+cot | 42.8 | 1.3 | 1.2 | 1.7 | 10 | 12.2 | 976 |
| magicoder-ds-7b | 41.7 | 1.5 | 0.89 | 1.7 | 10 | 11 | 972 |
| wizard-34b | 41.4 | 1.5 | 0.8 | 1.7 | 10 | 12.2 | 980 |
| deepseek-instruct-6.7b | 40.5 | 1.5 | 0.79 | 1.7 | 10 | 11.2 | 973 |
| codellama-python-34b | 39.7 | 1.5 | 0.82 | 1.7 | 10 | 10.8 | 972 |
| codellama-34b | 39.3 | 1.5 | 0.92 | 1.7 | 10 | 10.1 | 969 |
| phind | 38.9 | 1.5 | 0.85 | 1.7 | 10 | 10.1 | 966 |
| deepseek-base-6.7b | 38.3 | 1.4 | 0.92 | 1.7 | 10 | 9.53 | 965 |
| wizard-13b | 36.9 | 1.5 | 0.88 | 1.7 | 10 | 8.32 | 958 |
| codellama-python-13b | 36.4 | 1.4 | 0.91 | 1.7 | 10 | 8.51 | 958 |
| mixtral-8x7b | 36.3 | 1.4 | 0.94 | 1.7 | 10 | 8.35 | 956 |
| codellama-13b | 36.1 | 1.4 | 0.93 | 1.7 | 10 | 7.52 | 953 |
| codellama-13b+cot | 34.9 | 1.2 | 1.2 | 1.7 | 10 | 7.43 | 944 |
| codellama-python-7b | 32.4 | 1.4 | 0.9 | 1.7 | 10 | 6.43 | 942 |
| codellama-7b | 30.9 | 1.3 | 0.93 | 1.6 | 10 | 5.44 | 938 |
| starcoderbase-16b | 30.7 | 1.4 | 0.89 | 1.6 | 10 | 5.25 | 933 |
| mistral-7b | 30.1 | 1.3 | 0.96 | 1.6 | 10 | 5.11 | 931 |
| phi-2 | 29.7 | 1.3 | 0.92 | 1.6 | 10 | 5.46 | 930 |
| codellama-7b+cot | 29.1 | 1.1 | 1.1 | 1.6 | 10 | 4.09 | 920 |
| starcoderbase-7b | 28.9 | 1.3 | 0.88 | 1.6 | 10 | 4.95 | 930 |
| deepseek-instruct-1.3b | 27.4 | 1.4 | 0.79 | 1.6 | 10 | 5.58 | 925 |
| deepseek-base-1.3b | 25.9 | 1.2 | 0.94 | 1.5 | 10 | 4.26 | 917 |
| phi-1.5 | 21.7 | 1.1 | 0.91 | 1.5 | 10 | 3.82 | 898 |
| phi-1 | 19.3 | 1.2 | 0.78 | 1.4 | 10 | 4.19 | 896 |