List your company
ResearchFormerly WizWand

General Knowledge Evaluation on MMMLU (Accuracy)

82.25MMMLU General Knowledge AccuracyDOS-CPT
22.574838.067453.5669.0526Jun 1, 2025Jul 12, 2025Aug 22, 2025Oct 2, 2025Nov 12, 2025Dec 23, 2025Feb 2, 2026
Updated 8mo ago

Evaluation results

MethodLinks
DOS-CPTBackbone=Qwen3-14B-Base, CPT Strategy=Distance-to-Optimum Selection2025.1282.25—
RS-CPTBackbone=Qwen3-14B-Base, CPT Strategy=Random Data Sampling2025.1282.23—
HPS-CPTBackbone=Qwen3-14B-Base, CPT Strategy=High-PPL Sampling2025.1282.11—
Qwen3-14B-BaseBackbone=Qwen3-14B-Base, CPT Strategy=None2025.1282.03—
LPS-CPTBackbone=Qwen3-14B-Base, CPT Strategy=Low-PPL Sampling2025.1282.03—
Base modelBackbone=Llama3.2-3B Instruct2026.0260.45—
GA + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.0260.18—
NPO + KLRetention data (w/ Dr)=true, Backbone=Llama3.2-3B Instruct2026.0259.47—
NPOBackbone=Llama3.2-3B Instruct2026.0254.79—
CATNIPBackbone=Llama3.2-3B Instruct2026.0252.17—
SimNPOBackbone=Llama3.2-3B Instruct2026.0251.84—
FLATContrastive data (w/ Dct)=true, Backbone=Llama3.2-3B Instruct2026.0250.12—
ML-SFTModel Backbone=Qwen2.5-7B2025.0650.05—
CC-TUNINGModel Backbone=Qwen2.5-7B2025.0649.65—
CC-TUNING+MTModel Backbone=Qwen2.5-7B2025.0649.39—
ML-SFT+ENModel Backbone=Qwen2.5-7B2025.0648.8—
CC-TUNING+SDRRLModel Backbone=Qwen2.5-7B2025.0647.87—
ML-SFT+MTModel Backbone=Qwen2.5-7B2025.0647.49—
CC-TUNING+ENModel Backbone=Qwen2.5-7B2025.0647.35—
ML-SFT+SDRRLModel Backbone=Qwen2.5-7B2025.0647.28—
CC-TUNINGModel Backbone=LLaMA-3.1-8B2025.0640.74—
ML-SFT+MTModel Backbone=LLaMA-3.1-8B2025.0640.68—
ML-SFTModel Backbone=LLaMA-3.1-8B2025.0640.2—
ML-SFT+ENModel Backbone=LLaMA-3.1-8B2025.0639.62—
CC-TUNING+MTModel Backbone=LLaMA-3.1-8B2025.0638.87—
CC-TUNING+ENModel Backbone=LLaMA-3.1-8B2025.0638.73—
CC-TUNING+SDRRLModel Backbone=LLaMA-3.1-8B2025.0637.77—
ML-SFT+SDRRLModel Backbone=LLaMA-3.1-8B2025.0628.06—
GABackbone=Llama3.2-3B Instruct2026.0224.87—