Code Generation on HumanEval, MBPP, HumanEval+, MBPP+, CRUX-O
71.95HumanEval ScoreHSA-UL-Inst
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| HSA-UL-InstArchitecture=MoE, Total Parameters=8B, Activated Parameters=1B, Training Tokens=8T, Supervised fine-tuning=true2025.11 | 71.95 | 57 | 70.73 | 65.87 | 50.75 | |
| Qwen3-InstArchitecture=Dense, Total Parameters=1.7B, Activated Parameters=1.7B, Training Tokens=36T, Supervised fine-tuning=true2025.11 | 65.24 | 51 | 61.59 | 59.52 | 50 | |
| Qwen3-InstArchitecture=Dense, Total Parameters=0.6B, Activated Parameters=0.6B, Training Tokens=36T, Supervised fine-tuning=true2025.11 | 40.24 | 29.2 | 35.37 | 34.39 | 28 | |
| HSA-UL-InstArchitecture=Dense, Total Parameters=0.5B, Activated Parameters=0.5B, Training Tokens=4T, Supervised fine-tuning=true2025.11 | 39.63 | 34.4 | 37.2 | 39.95 | 23.25 |