Large Language Model Evaluation on MMLU, GSM8K, GPQA, HumanEval, TruthfulQA, and IFEVAL
70.7MMLUGRPO
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GRPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 70.7 | 75.7 | 28 | 64.3 | 59.9 | — | 62.1 | |
| TI-DPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 70 | 73 | 26 | 67 | 62 | — | 62.3 | |
| TIS-DPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 69.3 | 70.5 | 24.5 | 65.5 | 62.5 | — | 61.1 | |
| TPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 68.3 | 72.7 | 27.7 | 63.7 | 59 | — | 60.7 | |
| GRPOBackbone=Mistral-7B-v0.32025.05 | 68 | 52 | 9 | 51 | 64 | 56 | 50 | |
| CPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 67.3 | 70.7 | 26 | 62.8 | 58.3 | — | 59.4 | |
| KTOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 66.3 | 70.3 | 25.3 | 62 | 57.7 | — | 58.7 | |
| cDPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 66.1 | 70.1 | 25.1 | 61.9 | 57.6 | — | 58.5 | |
| TI-DPOBackbone=Mistral-7B-v0.32025.05 | 66 | 47 | 7 | 53 | 66 | 59 | 49.7 | |
| DPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 65.3 | 69.3 | 24 | 61 | 56.7 | — | 57.7 | |
| TPOBackbone=Mistral-7B-v0.32025.05 | 65 | 48 | 8 | 50 | 62 | 53 | 47.7 | |
| TDPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 65 | 68.2 | 23.5 | 60.3 | 56.3 | — | 57 | |
| CPOBackbone=Mistral-7B-v0.32025.05 | 64 | 46 | 7.5 | 49 | 61.5 | 51 | 46.5 | |
| SFTevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 64 | 68 | 22.7 | 59.3 | 55.5 | — | 56.7 | |
| Logic-RLevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 63.8 | 73.8 | 23.7 | 61 | 55.6 | — | 57.9 | |
| SimPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 63.5 | 64.7 | 21.8 | 58.2 | 54.2 | — | 54.5 | |
| KTOBackbone=Mistral-7B-v0.32025.05 | 63 | 45 | 7 | 48 | 61 | 50 | 45.7 | |
| IPOevaluation_setup=Average across three base models (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3)2025.05 | 63 | 65.3 | 20.3 | 57.3 | 52.7 | — | 54.2 | |
| DPOBackbone=Mistral-7B-v0.32025.05 | 62 | 44 | 6 | 47 | 60 | 50 | 44.8 | |
| TDPOBackbone=Mistral-7B-v0.32025.05 | 61 | 43 | 5.5 | 46 | 60 | 48 | 43.9 | |
| SFTBackbone=Mistral-7B-v0.32025.05 | 60 | 42 | 5 | 45 | 59.5 | 54 | 44.2 | |
| IPOBackbone=Mistral-7B-v0.32025.05 | 59 | 40 | 3 | 43 | 56 | 47 | 41.3 | |
| SIMPOBackbone=Mistral-7B-v0.32025.05 | 58 | 38 | 4 | 42 | 57 | 45 | 40.7 |