List your company
ResearchFormerly WizWand

Machine Translation (BLEU) on Flores

40.91BLEU ScoreBefore RL
14.431621.305828.1835.0542Jul 20, 2025Sep 7, 2025Oct 26, 2025Dec 15, 2025Feb 2, 2026Mar 23, 2026May 12, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Before RLModel=Qwen3-30B-A3B2026.0540.91
Vanilla RLModel=Qwen3-30B-A3B2026.0540.87
RL+SAEModel=Qwen3-30B-A3B2026.0540.84
RL+SAEModel=Qwen3-8B2026.0538.74
Vanilla RLModel=Qwen3-8B2026.0538.32
Before RLModel=Qwen3-8B2026.0538.18
SFTModel=Gemma-2-9B, Dataset=Russian 110k2025.0731.82
SFT+GRPOModel=Gemma-2-9B, Dataset=Russian 110k2025.0731.55
SASFTModel=Gemma-2-9B, Dataset=Russian 110k2025.0731.31
RL+SAEModel=Qwen3-1.7B2026.0531.06
SFT+GRPOModel=Gemma-2-9B2025.0730.71
SFTModel=Gemma-2-9B2025.0730.59
SFT+PenaltyModel=Gemma-2-9B, Dataset=Russian 110k2025.0730.32
SFT+GRPOModel=Qwen3-8B-Base2025.0730.14
SFT+GRPOModel=Qwen3-8B-Base, Training Dataset=Chinese 110k2026.0530.14
SFT+PenaltyModel=Qwen3-8B-Base2025.0730.1
SFT+PenaltyModel=Qwen3-8B-Base, Training Dataset=Chinese 110k2026.0530.1
SFT+PenaltyModel=Gemma-2-9B2025.0730.09
SFTModel=Qwen3-8B-Base2025.0729.99
SFTModel=Qwen3-8B-Base, Training Dataset=Chinese 110k2026.0529.99
SASFTModel=Qwen3-8B-Base2025.0729.97
SASFTModel=Qwen3-8B-Base, Training Dataset=Chinese 110k2026.0529.97
SFT+GRPOModel=Gemma-2-9B2025.0729.68
Vanilla RLModel=Qwen3-1.7B2026.0529.44
SASFTModel=Gemma-2-9B2025.0729.41
SFT+PenaltyBackbone=Gemma-2-9B, Training dataset setting=Korean 210k2025.0729.29
SFTModel=Gemma-2-9B2025.0729.23
SFT+PenaltyModel=Gemma-2-9B2025.0729.18
SFTBackbone=Gemma-2-9B, Training dataset setting=Korean 210k2025.0729.15
SASFTBackbone=Gemma-2-9B, Training dataset setting=Korean 210k2025.0728.87
SFT+GRPOBackbone=Gemma-2-9B, Training dataset setting=Korean 210k2025.0728.68
Before RLModel=Qwen3-1.7B2026.0528.47
SASFTModel=Gemma-2-9B2025.0728.19
SFT+GRPOBackbone=Qwen3-8B-Base, Training dataset setting=Korean 210k2025.0727.07
SASFTModel=Qwen3-8B-Base, Dataset=Russian 110k2025.0726.95
SFT+PenaltyModel=Qwen3-8B-Base, Dataset=Russian 110k2025.0726.89
SFT+GRPOModel=Qwen3-8B-Base, Dataset=Russian 110k2025.0726.86
SFTBackbone=Qwen3-8B-Base, Training dataset setting=Korean 210k2025.0726.37
SFT+PenaltyModel=Qwen3-8B-Base2025.0726.12
SFTModel=Qwen3-8B-Base, Dataset=Russian 110k2025.0725.77
SFT+PenaltyModel=Llama-3.1-8B, Dataset=Russian 110k2025.0725.49
SFT+PenaltyBackbone=Qwen3-8B-Base, Training dataset setting=Korean 210k2025.0724.83
SFT+GRPOModel=Qwen3-8B-Base2025.0724.81
SFTModel=Llama-3.1-8B2025.0724.73
SFT+PenaltyModel=Llama-3.1-8B2025.0724.68
SASFTModel=Qwen3-1.7B-Base, Dataset=Russian 110k2025.0724.56
SFT+PenaltyModel=Llama-3.1-8B2025.0724.36
SASFTModel=Llama-3.1-8B2025.0724.05
SFT+GRPOModel=Qwen3-1.7B-Base, Dataset=Russian 110k2025.0723.84
SFT+GRPOBackbone=Llama-3.1-8B, Training dataset setting=Korean 210k2025.0723.79
SASFTModel=Llama-3.1-8B2025.0723.73
SFTModel=Qwen3-1.7B-Base2025.0723.7
SFTModel=Qwen3-1.7B-Base, Training Dataset=Chinese 110k2026.0523.7
SFTBackbone=Llama-3.1-8B, Training dataset setting=Korean 210k2025.0723.68
SFT+PenaltyModel=Qwen3-1.7B-Base, Dataset=Russian 110k2025.0723.68
SASFTModel=Qwen3-1.7B-Base2025.0723.67
SASFTModel=Qwen3-1.7B-Base, Training Dataset=Chinese 110k2026.0523.67
SFT+PenaltyModel=Qwen3-1.7B-Base2025.0723.55
SFT+PenaltyModel=Qwen3-1.7B-Base, Training Dataset=Chinese 110k2026.0523.55
SFT+GRPOModel=Gemma-2-2B, Dataset=Russian 110k2025.0723.54
SASFTModel=Llama-3.1-8B, Dataset=Russian 110k2025.0723.52
SFTModel=Qwen3-1.7B-Base, Dataset=Russian 110k2025.0723.46
SFT+GRPOModel=Qwen3-1.7B-Base2025.0723.45
SFT+GRPOModel=Qwen3-1.7B-Base, Training Dataset=Chinese 110k2026.0523.45
SFT+PenaltyBackbone=Llama-3.1-8B, Training dataset setting=Korean 210k2025.0723.44
SASFTModel=Gemma-2-2B, Dataset=Russian 110k2025.0723.31
SASFTBackbone=Llama-3.1-8B, Training dataset setting=Korean 210k2025.0723.25
SFTModel=Gemma-2-2B, Dataset=Russian 110k2025.0723.25
SFT+GRPOModel=Qwen3-1.7B-Base2025.0723.04
SASFTModel=Qwen3-1.7B-Base2025.0722.97
SFT+PenaltyModel=Gemma-2-2B2025.0722.87
SFTModel=Qwen3-8B-Base2025.0722.86
SFTModel=Llama-3.1-8B2025.0722.81
SFT+GRPOModel=Gemma-2-2B2025.0722.8
SFT+PenaltyModel=Qwen3-1.7B-Base2025.0722.58
SFTModel=Gemma-2-2B2025.0722.56
SASFTBackbone=Qwen3-1.7B-Base, Training dataset setting=Korean 210k2025.0722.55
SFT+PenaltyBackbone=Qwen3-1.7B-Base, Training dataset setting=Korean 210k2025.0722.5
SFTModel=Qwen3-1.7B-Base2025.0722.49
SFT+GRPOBackbone=Qwen3-1.7B-Base, Training dataset setting=Korean 210k2025.0722.44
SFTBackbone=Qwen3-1.7B-Base, Training dataset setting=Korean 210k2025.0722.4
SFT+GRPOModel=Llama-3.1-8B2025.0722.34
SASFTModel=Gemma-2-2B2025.0722.28
SFT+PenaltyModel=Gemma-2-2B, Dataset=Russian 110k2025.0722.18
SFT+GRPOModel=Llama-3.1-8B, Dataset=Russian 110k2025.0721.82
SASFTBackbone=Qwen3-8B-Base, Training dataset setting=Korean 210k2025.0721.8
SFTModel=Llama-3.1-8B, Dataset=Russian 110k2025.0721.45
SFT+GRPOModel=Llama-3.1-8B2025.0720.75
SFT+GRPOBackbone=Gemma-2-2B, Training dataset setting=Korean 210k2025.0719.35
SFTBackbone=Gemma-2-2B, Training dataset setting=Korean 210k2025.0719.31
SASFTModel=Qwen3-8B-Base2025.0718.96
SFT+GRPOModel=Gemma-2-2B2025.0718.49
SFT+PenaltyModel=Gemma-2-2B2025.0718.48
SFTModel=Gemma-2-2B2025.0718.39
SASFTBackbone=Gemma-2-2B, Training dataset setting=Korean 210k2025.0718.34
SASFTModel=Gemma-2-2B2025.0717.96
SFT+PenaltyBackbone=Gemma-2-2B, Training dataset setting=Korean 210k2025.0715.45