List your company
ResearchFormerly WizWand

Code Generation on CodeContests (Avg@8)

39.33Avg@8CodeScaler
18.21823.69929.1834.661Feb 4, 2026Feb 5, 2026Feb 7, 2026Feb 9, 2026Feb 11, 2026Feb 13, 2026Feb 15, 2026
Updated 8mo ago

Evaluation results

MethodLinks
CodeScalerBase Model=Qwen3-14B-Base, Reward Feedback Source=CodeScaler2026.0239.33
RLVR (Binary)Base Model=Qwen3-14B-Base, Reward Feedback Source=RLVR (Binary)2026.0234.3
CodeScalerBase Model=Qwen3-8B-Base, Reward Feedback Source=CodeScaler2026.0233.94
RLVR (Binary)Training Dataset=rStarCoder, Backbone Model=Qwen3-14B, Training Protocol=RLVR2026.0233.26
CodeScalerTraining Dataset=KodCode, Backbone Model=Qwen3-14B, Training Protocol=CodeScaler2026.0232.89
CodeScalerTraining Dataset=rStarCoder, Backbone Model=Qwen3-14B, Training Protocol=CodeScaler2026.0232.63
RLVR (Binary)Training Dataset=KodCode, Backbone Model=Qwen3-14B, Training Protocol=RLVR2026.0232.32
AceCodeRMBase Model=Qwen3-14B-Base, Reward Feedback Source=AceCodeRM2026.0232.11
AEROBackbone=Qwen2.5-7B-Instruct-1M, Time (min)=6.95, FLOPs (PetaFLOP)=2295, p-value=—2026.0232.1
RLVR (Binary)Base Model=Qwen3-8B-Base, Reward Feedback Source=RLVR (Binary)2026.0232
SkyworkRMBase Model=Qwen3-14B-Base, Reward Feedback Source=SkyworkRM2026.0231.64
DAPOBackbone=Qwen2.5-7B-Instruct-1M, Time (min)=7.75, FLOPs (PetaFLOP)=2941, p-value== 0.0342026.0231.3
CodeScalerTraining Dataset=rStarCoder, Backbone Model=Qwen3-8B, Training Protocol=CodeScaler2026.0230.75
GRPOBackbone=Qwen2.5-7B-Instruct-1M, Time (min)=11.66, FLOPs (PetaFLOP)=4488, p-value=< 0.0052026.0229.6
RLVR (Binary)Training Dataset=rStarCoder, Backbone Model=Qwen3-8B, Training Protocol=RLVR2026.0229.44
CodeScalerTraining Dataset=KodCode, Backbone Model=Qwen3-8B, Training Protocol=CodeScaler2026.0229.34
AceCodeRMBase Model=Qwen3-8B-Base, Reward Feedback Source=AceCodeRM2026.0228.34
RLVR (Binary)Training Dataset=KodCode, Backbone Model=Qwen3-8B, Training Protocol=RLVR2026.0227.87
Qwen3-14B-BaseTraining Dataset=KodCode, Backbone Model=Qwen3-14B, Training Protocol=Base2026.0226.25
Qwen3-14B-BaseTraining Dataset=rStarCoder, Backbone Model=Qwen3-14B, Training Protocol=Base2026.0226.25
Qwen3-14B-BaseBase Model=Qwen3-14B-Base, Reward Feedback Source=None (Base)2026.0226.25
SkyworkRMBase Model=Qwen3-8B-Base, Reward Feedback Source=SkyworkRM2026.0223.22
Base ModelBackbone=Qwen2.5-7B-Instruct-1M, Time (min)=—, FLOPs (PetaFLOP)=—, p-value=< 1e-52026.0219.5
Qwen3-8B-BaseTraining Dataset=KodCode, Backbone Model=Qwen3-8B, Training Protocol=Base2026.0219.03
Qwen3-8B-BaseTraining Dataset=rStarCoder, Backbone Model=Qwen3-8B, Training Protocol=Base2026.0219.03
Qwen3-8B-BaseBase Model=Qwen3-8B-Base, Reward Feedback Source=None (Base)2026.0219.03