Summarization on XLSum Japanese (test)
12.95ROUGE-2Llama3 8B-I + LATB
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Llama3 8B-I + LATBAlignment Method=Language-Aware Token Boosting2026.06 | 12.95 | 17.7 | — | — | 27.54 | |
| Llama3 8B-I + Adaptive LATBAlignment Method=Adaptive Language-Aware Token Boosting2026.06 | 12.92 | 17.89 | — | — | 27.89 | |
| Llama3 8B-IPrompting Strategy=Default2026.06 | 12.53 | 16.84 | — | — | 26.42 | |
| Llama3 8B-IPrompting Strategy=Strict Prompt2026.06 | 12.43 | 16.97 | — | — | 26.48 | |
| Suzume 8B-Multilingual2026.06 | 11.73 | 16.55 | — | — | 26.13 | |
| Spec.shots=0, draft_model=Llama-2 7B, draft_tokens=52024.01 | 11.6 | 26.5 | 28.5 | 0.97 | — | |
| MuMoshots=02024.01 | 11.6 | 26.3 | 59.2 | 2.02 | — | |
| Shortlistingshots=0, target_language_subset=mC42024.01 | 11.4 | 26.3 | 30.3 | 1.03 | — | |
| Vanilla Decodingshots=02024.01 | 11.3 | 26.6 | 29.3 | 1 | — | |
| Spec. (w/o Rejection)shots=0, draft_model=Llama-2 7B, draft_tokens=5, rejection_module=false2024.01 | 10.8 | 24.2 | 35.4 | 1.21 | — |