Code Generation on HumanEval (Accuracy, Loss)
47AccuracyAdam+Nexus
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Adam+NexusOptimizer=Adam+Nexus, Model Size=3B2026.04 | 47 | 1.086 | |
| AdamWOptimizer=AdamW, Model Size=3B2026.04 | 43 | 1.116 | |
| MuonOptimizer=Muon, Model Size=3B2026.04 | 40 | 1.091 |