Functionally Diverse Response Generation on Functional Diversity Prompt Set Categories A-H 1.0 (test)
4.26Score Category AGRPO w/DARLING
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| GRPO w/DARLINGTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 4.26 | 3.19 | 3.64 | 2.65 | 2.52 | 2.13 | 3.53 | 1.89 | |
| GRPOTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 4 | 2.69 | 3.93 | 3.09 | 1.16 | 1.04 | 1.67 | 1.36 | |
| GRPOTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 3.94 | 2.75 | 4.29 | 2.67 | 1.5 | 1.83 | 2.58 | 1.51 | |
| GRPO w/DARLINGTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.94 | 2.25 | 3.29 | 2.75 | 2.06 | 2.22 | 2.82 | 1.64 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 3.89 | 2.25 | 3.21 | 3.47 | 4.46 | 4 | 4.36 | 3.32 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Ultrachat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 3.68 | 2.31 | 3.14 | 2.65 | 3.14 | 3.17 | 3.84 | 2.43 | |
| GRPOTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.66 | 2.56 | 3.64 | 2.73 | 1.4 | 1.87 | 2.36 | 1.45 | |
| GRPOTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.55 | 2.31 | 3.29 | 3.13 | 1.18 | 1.17 | 1.47 | 1.27 | |
| GRPO w/DARLINGTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=1.0)2025.09 | 3.53 | 1.75 | 3.14 | 2.25 | 1.52 | 1.83 | 2.16 | 1.33 | |
| GRPO w/DARLINGTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 3.47 | 1.69 | 3.14 | 2.6 | 1.7 | 1.83 | 2.11 | 1.35 | |
| GRPO w/DARLINGTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.42 | 1.56 | 2.86 | 2.38 | 1.58 | 1.83 | 2.13 | 1.28 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.42 | 1.88 | 2.43 | 3.29 | 3.98 | 3.43 | 3.93 | 2.92 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Ultrachat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.5)2025.09 | 3.23 | 1.62 | 2.07 | 2.44 | 2.66 | 2.87 | 2.89 | 1.81 | |
| GRPOTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 3.15 | 1.69 | 3.14 | 2.25 | 1.2 | 1.52 | 1.96 | 1.24 | |
| GRPOTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 2.79 | 1.62 | 2.71 | 2.84 | 1.12 | 1.09 | 1.22 | 1.12 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Wildchat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 2.58 | 1.19 | 2.21 | 3.07 | 3.1 | 2.7 | 3.44 | 2.11 | |
| GRPO w/DARLINGTraining Dataset=Ultrafeedback, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 2.43 | 1.19 | 2.29 | 2.29 | 1.5 | 1.26 | 1.62 | 1.1 | |
| GRPO w/DARLING & Task Diversity JudgesTraining Dataset=Ultrachat, beta=0.001, steps=1000, Sampling Strategy=Temperature (t=0.1)2025.09 | 1.87 | 1.25 | 1.21 | 1.95 | 1.78 | 2 | 2.22 | 1.33 | |
| Llama-3.1-8B-InstructSampling Strategy=Temperature (t=1.0)2025.09 | 1.45 | 2.31 | 3.71 | 2.42 | 2.68 | 2 | 2.51 | 1.78 | |
| Llama-3.1-8B-InstructSampling Strategy=Temperature (t=0.5)2025.09 | 1.19 | 1.75 | 2.79 | 2.22 | 2.02 | 1.48 | 1.91 | 1.47 | |
| Llama-3.1-8B-InstructSampling Strategy=Temperature (t=0.1)2025.09 | 1.08 | 1.25 | 1.71 | 1.93 | 1.56 | 1.52 | 1.47 | 1.09 |