Knowledge Acquisition on Wikipedia (test) (In-distribution)
91Accuracy (strict)Oracle RAG
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Oracle RAGBackbone=Qwen2.5-7B-Instruct2026.01 | 91 | 100 | |
| SDFTBackbone=Qwen2.5-7B-Instruct2026.01 | 89 | 100 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.01 | 80 | 95 | |
| CPTBackbone=Qwen2.5-7B-Instruct2026.01 | 9 | 37 | |
| BaseBackbone=Qwen2.5-7B-Instruct2026.01 | 0 | 0 |