List your company

Three-objective alignment on Helpful Assistant (test)

0.74Helpfulness ScoreRLHF-r1
-0.9968-0.5459-0.0950.3559May 16, 2025

Evaluation results

MethodLinks
RLHF-r1SFT model=Mistral-7b-v0.2 (Instruct)2025.050.74-0.41-0.22
RLHF-r1SFT model=OpenChat-v3.52025.050.72-0.38-0.2
RLHF-r1SFT model=Llama-3.1-8B2025.050.69-0.33-0.18
RLHF-r1SFT model=phi-1.52025.050.65-0.73-0.47
PARMSFT model=Mistral-7b-v0.2 (Instruct)2025.050.450.160.23
MOPO-LagSFT model=OpenChat-v3.52025.050.40.230.18
MOPO-LagSFT model=Mistral-7b-v0.2 (Instruct)2025.050.390.220.17
MOPO-LagSFT model=Llama-3.1-8B2025.050.380.210.16
PARMSFT model=OpenChat-v3.52025.050.320.180.24
MOPO-LBSFT model=Llama-3.1-8B2025.050.310.20.19
PARMSFT model=Llama-3.1-8B2025.050.30.170.22
MOPO-LBSFT model=Mistral-7b-v0.2 (Instruct)2025.050.30.190.18
MOPO-LBSFT model=OpenChat-v3.52025.050.290.180.17
RiCSFT model=OpenChat-v3.52025.050.270.160.12
RiCSFT model=Llama-3.1-8B2025.050.260.140.11
RiCSFT model=Mistral-7b-v0.2 (Instruct)2025.050.250.150.1
MOPO-LBSFT model=phi-1.52025.050.240.130.14
PARMSFT model=phi-1.52025.050.230.120.11
MOPO-LagSFT model=phi-1.52025.050.20.10.09
DPO on DJSFT model=Llama-3.1-8B2025.050.190.10.12
RiCSFT model=phi-1.52025.050.180.10.05
DPO on DJSFT model=Mistral-7b-v0.2 (Instruct)2025.050.180.090.11
DPO on DJSFT model=OpenChat-v3.52025.050.170.080.1
DPO on DJSFT model=phi-1.52025.050.130.050.08
MODPOSFT model=OpenChat-v3.52025.050.05-0.080.09
MODPOSFT model=Mistral-7b-v0.2 (Instruct)2025.050.04-0.090.08
MODPOSFT model=Llama-3.1-8B2025.050.03-0.10.07
MODPOSFT model=phi-1.52025.05-0.04-0.170
RLHF-r3SFT model=OpenChat-v3.52025.05-0.77-0.890.4
RLHF-r2SFT model=OpenChat-v3.52025.05-0.780.49-0.37
RLHF-r3SFT model=Mistral-7b-v0.2 (Instruct)2025.05-0.78-0.910.41
RLHF-r2SFT model=Mistral-7b-v0.2 (Instruct)2025.05-0.790.52-0.38
RLHF-r2SFT model=Llama-3.1-8B2025.05-0.80.5-0.39
RLHF-r3SFT model=Llama-3.1-8B2025.05-0.8-0.90.43
RLHF-r3SFT model=phi-1.52025.05-0.88-0.970.29
RLHF-r2SFT model=phi-1.52025.05-0.930.44-0.53