General Capability Retention on GSM8K, MMLU, and ARC-C
68.5GSM8KOGS
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OGSModel=Qwen3-8B, Data Selection Ratio=10%2026.02 | 68.5 | 72.8 | 74.9 | 72.1 | |
| GRADSModel=Qwen3-8B, Data Selection Ratio=10%2026.02 | 66.8 | 71.2 | 73.5 | 70.5 | |
| Influence Func.Model=Qwen3-8B, Data Selection Ratio=10%2026.02 | 65.2 | 69.8 | 72 | 69 | |
| RandomModel=Qwen3-8B, Data Selection Ratio=10%2026.02 | 65.1 | 70.2 | 72.1 | 69.1 | |
| PerplexityModel=Qwen3-8B, Data Selection Ratio=10%2026.02 | 64.5 | 69.5 | 71.5 | 68.5 | |
| LESSModel=Qwen3-8B, Data Selection Ratio=10%2026.02 | 60.1 | 65.4 | 68.2 | 64.6 | |
| Full DataModel=Qwen3-8B, Data Selection Ratio=100%2026.02 | 55.3 | 60.1 | 65.4 | 60.3 | |
| OGSModel=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 55.2 | 64.5 | 66.2 | 62 | |
| GRADSModel=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 53.5 | 63.1 | 64.8 | 60.5 | |
| Influence Func.Model=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 52 | 62.5 | 64 | 59.5 | |
| PerplexityModel=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 51.2 | 61.8 | 63.5 | 58.8 | |
| RandomModel=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 50.5 | 62.1 | 64.3 | 59 | |
| LESSModel=Llama-3.1-8B, Data Selection Ratio=10%2026.02 | 48.2 | 58.4 | 60.1 | 55.6 | |
| Full DataModel=Llama-3.1-8B, Data Selection Ratio=100%2026.02 | 45.1 | 55.2 | 58.9 | 53.1 |