StarVLA/Qwen3-VL-OFT-LIBERO-4in1

LIBERO Evaluation Results

Qwen3-VL-OFT-LIBERO-4in1 Evaluation Results

Steps	libero_object	libero_spatial	libero_goal	libero_10	Avg
20000	0.97	0.984	0.97	0.874	0.9495
30000	0.994	0.982	0.97	0.916	0.9655
40000	0.998	0.99	0.968	0.928	0.9710
50000	0.996	0.99	0.986	0.948	0.9800

📈 Convergence Trends

libero_object: Rapid convergence, reaching peak performance of 0.998 at 40k steps
libero_spatial: Stable high-level performance, all steps >0.98
libero_goal: Continuous improvement, achieving best performance 0.986 at 50k steps
libero_10: Most challenging task but consistently improving, from 0.874 to 0.948

🎯 Optimal Configuration

• Best step count: 50k steps (average success rate 98.00%)
• Most challenging task: libero_10 (average 91.65%)
• Easiest task: libero_object (average 98.95%)

🔍 Key Findings

1. Stable convergence: All datasets reached or approached optimal performance at 50k steps
2. Generalization capability: The model shows strong generalization ability, with excellent performance on 4-task joint training
3. Challenging tasks: libero_10 remains the most challenging task with room for improvement
4. Training efficiency: Stable performance between 40k-50k steps indicates no overfitting
5. Overall performance: Average success rate steadily increased from 94.95% at 20k steps to 98.00% at 50k steps