Abstractive Summarization on BillSum
59.67ROUGE-1PEGASUS_LARGE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| PEGASUS_LARGEPre-training strategy=mixed, stochastic2019.12 | 59.67 | 41.58 | 47.59 | |
| non-privateProj Dim. (k)=N/A, Epsilon (ϵ)=N/A, Fine-tuning protocol=full fine-tuning, Model=Llama-3.2-1B2026.05 | 49.28 | — | — | |
| DP-SGD (FGC)Proj Dim. (k)=N/A, Epsilon (ϵ)=9, Fine-tuning protocol=full fine-tuning, Model=Llama-3.2-1B2026.05 | 48.82 | — | — | |
| DP-SGD-RCProj Dim. (k)=32, Epsilon (ϵ)=9, Fine-tuning protocol=full fine-tuning, Model=Llama-3.2-1B2026.05 | 48.64 | — | — | |
| DP-SGD (FGC)Proj Dim. (k)=N/A, Epsilon (ϵ)=2, Fine-tuning protocol=full fine-tuning, Model=Llama-3.2-1B2026.05 | 48.31 | — | — | |
| DP-SGD-RCProj Dim. (k)=32, Epsilon (ϵ)=2, Fine-tuning protocol=full fine-tuning, Model=Llama-3.2-1B2026.05 | 47.96 | — | — |