Question Answering on Natural Questions (NQ) (EM and Performance Gain Metrics)
49.8Exact Match (EM)PatchRAG
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PatchRAGStep=t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 49.8 | 13.6 | |
| ChatQA-1.5Step=-, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 42.4 | — | |
| RAFTStep=t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 41.9 | 5.5 | |
| Golden-onlyStep=t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 40.2 | 4.9 | |
| Auto-RAGStep=-, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 37.9 | — | |
| Self-RAGStep=-, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 36.4 | — | |
| RAFTStep=< t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 36.4 | — | |
| PatchRAGStep=< t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 36.2 | — | |
| Golden-onlyStep=< t, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 35.3 | — | |
| Standard RAGStep=-, Generator=Llama-3 8B, Retriever=bge-m32026.04 | 28.7 | — |