List your company
ResearchFormerly WizWand

Alignment on MT-Bench

9.12MT-Bench ScoreQwen2-72B-Instruct
3.62885.05446.487.9056Jul 15, 2024Nov 1, 2024Feb 18, 2025Jun 7, 2025Sep 24, 2025Jan 11, 2026Apr 30, 2026
Updated 6mo ago

Evaluation results

MethodLinks
Qwen2-72B-InstructType=Instruction-tuned2024.079.12———
Llama-3-70B-InstructType=Instruction-tuned2024.078.95———
Qwen1.5-110B-ChatType=Instruction-tuned2024.078.88———
Mixtral-8x22B-InstructType=Instruction-tuned2024.078.66———
Qwen1.5-72B-ChatType=Instruction-tuned2024.078.61———
FsfairX + LPRM=+ LP, Backbone=Gemma-2-9B-it2026.048.16———
FsfairX + CIRMRM=+ CIRM, Backbone=Gemma-2-9B-it2026.048.15———
INFRM=INF, Backbone=Gemma-2-9B-it2026.048.11———
FsfairXRM=FsfairX, Backbone=Gemma-2-9B-it2026.048.07———
GRMRM=GRM, Backbone=Gemma-2-9B-it2026.048.01———
FsfairX + LWRRM=+ LWR, Backbone=Gemma-2-9B-it2026.047.96———
GRM + CIRMRM=+ CIRM, Backbone=Gemma-2-9B-it2026.047.89———
GRM + LWRRM=+ LWR, Backbone=Gemma-2-9B-it2026.047.87———
GRM + LPRM=+ LP, Backbone=Gemma-2-9B-it2026.047.84———
Gemma-2-9B-itRM=-, Backbone=Gemma-2-9B-it2026.047.83———
FsfairXRM=FsfairX, Backbone=Llama-3-8B-Instruct2026.047.64———
FsfairX + CIRMRM=+ CIRM, Backbone=Llama-3-8B-Instruct2026.047.62———
FsfairX + LPRM=+ LP, Backbone=Llama-3-8B-Instruct2026.047.6———
GRM + LWRRM=+ LWR, Backbone=Llama-3-8B-Instruct2026.047.58———
GRM + CIRMRM=+ CIRM, Backbone=Llama-3-8B-Instruct2026.047.53———
GRMRM=GRM, Backbone=Llama-3-8B-Instruct2026.047.45———
FsfairX + LWRRM=+ LWR, Backbone=Llama-3-8B-Instruct2026.047.44———
INFRM=INF, Backbone=Llama-3-8B-Instruct2026.047.42———
Llama-3-8B-InstructRM=-, Backbone=Llama-3-8B-Instruct2026.047.34———
GRM + LPRM=+ LP, Backbone=Llama-3-8B-Instruct2026.047.29———
FullBackbone=LLaMA-3-8B, % Train=100%, Alignment Method=DPO2025.055.51——49.4
FullBackbone=LLaMA-3-8B, % Train=100%, Alignment Method=SimPO2025.055.42——47.5
RandomBackbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.055.28——47.5
HighVar.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.055.21——43.1
HighAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.055.19——48.1
LowAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.055.15——48.8
FullBackbone=Mistral-7B, % Train=100%, Alignment Method=DPO2025.055.15——49.7
HighAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.055.13——50.9
LowAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.055.01——45
RandomBackbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.054.98——45
HighAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.054.96——45.6
HighVar.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.054.8——38.4
ZeroshotBackbone=LLaMA-3-8B, % Train=0%, Alignment Method=DPO2025.054.61——27.8
ZeroshotBackbone=LLaMA-3-8B, % Train=0%, Alignment Method=SimPO2025.054.61——27.8
LowAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.054.61——33.1
RandomBackbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.054.54——33.1
HighAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.054.51——34.1
HighVar.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.054.49——30
FullBackbone=Mistral-7B, % Train=100%, Alignment Method=SimPO2025.054.46——34.7
HighVar.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.054.4——23.1
ZeroshotBackbone=Mistral-7B, % Train=0%, Alignment Method=DPO2025.054.04——21.6
ZeroshotBackbone=Mistral-7B, % Train=0%, Alignment Method=SimPO2025.054.04——21.6
LowAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.053.88——19.4
RandomBackbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.053.84——20.3
PSFTBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback2025.08—6.834.86—
PSFT → DPOBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback, Post-training=DPO2025.08—8.516.95—
PSFTprolongBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback2025.08—7.415.84—
PSFTprolong → DPOBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback, Post-training=DPO2025.08—7.636.74—
SFTBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback2025.08—7.645.71—
SFT → DPOBackbone=Qwen3-4B-Base, Training Dataset=UltraFeedback, Post-training=DPO2025.08—7.916—