Evaluating Context Influence and Input Regurgitation on PubMedQA
97.95Context Influence Score I(D; y_tilde)OPT 1.3B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OPT 1.3BDecoding λ=1.52024.10 | 97.95 | 151 | 494 | |
| GPT-Neo 1.3BDecoding λ=1.52024.10 | 77.91 | 206 | 622 | |
| LLaMA 3 8BDecoding λ=1.52024.10 | 70.91 | 123 | 669 | |
| OPT 1.3BDecoding λ=1.02024.10 | 45.66 | 47 | 251 | |
| GPT-Neo 1.3BDecoding λ=1.02024.10 | 38.79 | 54 | 268 | |
| LLaMA 3 8BDecoding λ=1.02024.10 | 37.01 | 58 | 439 | |
| LLaMA 3 8BDecoding λ=0.52024.10 | 16.69 | 0 | 128 | |
| OPT 1.3BDecoding λ=0.52024.10 | 13.2 | 0 | 54 | |
| GPT-Neo 1.3BDecoding λ=0.52024.10 | 11.2 | 0 | 53 |