Data-to-Text Generation on DART (test)
52BLEUC-P (large)
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| C-P (large)Linear=O, Training Protocol=Multi-task2023.08 | 52 | — | — | — | — | — | — | — | — | |
| T5-3BLinear=Uni, Training Protocol=Multi-task2023.08 | 50.2 | — | — | — | — | — | — | — | — | |
| T5-baseLinear=Uni, Training Protocol=Multi-task2023.08 | 49.8 | — | — | — | — | — | — | — | — | |
| T5-3BLinear=O, Training Protocol=Multi-task2023.08 | 49.8 | — | — | — | — | — | — | — | — | |
| T5-3BLinear=Uni, Training Protocol=Single-task2023.08 | 49.6 | — | — | — | — | — | — | — | — | |
| T5-3BLinear=O, Training Protocol=Single-task2023.08 | 49.3 | — | — | — | — | — | — | — | — | |
| T5-basesource=Nan et al. [12], training_data=Full training set2024.01 | 49.21 | 40 | 44 | — | — | — | — | — | — | |
| T5-baseLinear=O, Training Protocol=Single-task2023.08 | 49 | — | — | — | — | — | — | — | — | |
| T5-baseLinear=Uni, Training Protocol=Single-task2023.08 | 48.6 | — | — | — | — | — | — | — | — | |
| T5-basesource=Our run, training_data=Full training set2024.01 | 48.52 | 39.7 | 45.91 | — | — | — | — | — | — | |
| T5-baseLinear=O, Training Protocol=Multi-task2023.08 | 48.1 | — | — | — | — | — | — | — | — | |
| self-mem + new datadata_subset=30% fixed data per epoch2024.01 | 47.76 | 39.51 | 48.28 | — | — | — | — | — | — | |
| self-mem + new data + self T2Ddata_subset=30% fixed data per epoch2024.01 | 47.6 | 39.39 | 48.23 | — | — | — | — | — | — | |
| self-mem + new datadata_subset=30% random data per epoch2024.01 | 47.54 | 39.38 | 48.07 | — | — | — | — | — | — | |
| LoRABackbone=GPT-2 Large, # Trainable Parameters=0.77M2021.06 | 47.5 | 39 | 45 | — | — | — | — | — | — | |
| self-mem + new data + self T2Ddata_subset=30% random data per epoch2024.01 | 47.32 | 39.19 | 47.34 | — | — | — | — | — | — | |
| BART-basesource=Nan et al. [12], training_data=Full training set2024.01 | 47.11 | 38 | 46 | — | — | — | — | — | — | |
| LoRABackbone=GPT-2 Medium, # Trainable Parameters=0.35M2021.06 | 47.1 | 39 | 46 | — | — | — | — | — | — | |
| Adapter (H)Backbone=GPT-2 Large, # Trainable Parameters=23M2021.06 | 47.1 | 39 | 45 | — | — | — | — | — | — | |
| Fine-TuneBackbone=GPT-2 Large, # Trainable Parameters=774M2021.06 | 47 | 39 | 46 | — | — | — | — | — | — | |
| no self-mem 3data_subset=30% random data per epoch2024.01 | 46.9 | 38.92 | 46.61 | — | — | — | — | — | — | |
| PrefLayerBackbone=GPT-2 Large, # Trainable Parameters=0.77M2021.06 | 46.7 | 38 | 45 | — | — | — | — | — | — | |
| UnifiedSKG (3B)Linear=O, Training Protocol=Single-task2023.08 | 46.7 | — | — | — | — | — | — | — | — | |
| PrefLayerBackbone=GPT-2 Medium, # Trainable Parameters=0.35M2021.06 | 46.4 | 38 | 46 | — | — | — | — | — | — | |
| Fine-TuneBackbone=GPT-2 Medium, # Trainable Parameters=354M2021.06 | 46.2 | 39 | 46 | — | — | — | — | — | — | |
| UnifiedSKG (base)Linear=O, Training Protocol=Single-task2023.08 | 46.2 | — | — | — | — | — | — | — | — | |
| BART-basesource=Our run, training_data=Full training set2024.01 | 46.15 | 38.13 | 48.2 | — | — | — | — | — | — | |
| self-mem + self T2Ddata_subset=30% fixed data per epoch2024.01 | 46.14 | 38.86 | 49.08 | — | — | — | — | — | — | |
| Adapter (L)Backbone=GPT-2 Large, # Trainable Parameters=0.88M2021.06 | 45.7 | 38 | 46 | — | — | — | — | — | — | |
| self-memdata_subset=30% fixed data per epoch2024.01 | 45.63 | 38.9 | 49.1 | — | — | — | — | — | — | |
| self-mem + self T2Ddata_subset=30% random data per epoch2024.01 | 45.62 | 38.81 | 48.95 | — | — | — | — | — | — | |
| FLAN-T5-basesource=Our run, training_data=Full training set2024.01 | 45.53 | 37.36 | 48.54 | — | — | — | — | — | — | |
| AdaPreLoRA AdamWModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 45.4 | 67 | 0.49 | — | 60 | 40 | — | — | — | |
| LoRABackbone=GPT-2 medium, # Params (M)=0.3M, Beam size=102025.06 | 45.35 | 38 | 53 | — | — | — | — | — | — | |
| Adapter (H)Backbone=GPT-2 Medium, # Trainable Parameters=11M2021.06 | 45.2 | 38 | 46 | — | — | — | — | — | — | |
| LoRA-Pro AdamWModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 44.9 | 66 | 0.5 | — | 62 | 39 | — | — | — | |
| no self-mem 1data_subset=30% fixed data per epoch2024.01 | 44.86 | 37.98 | 50.83 | — | — | — | — | — | — | |
| Scaled AdamWModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 44.8 | 67 | 0.49 | — | 62 | 40 | — | — | — | |
| AdaPreLoRA SGDModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 44.6 | 66 | 0.49 | — | 62 | 39 | — | — | — | |
| self-memdata_subset=30% random data per epoch2024.01 | 44.52 | 38.62 | 50.75 | — | — | — | — | — | — | |
| LoRA-Pro SGDModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 44.1 | 66 | 0.5 | — | 61 | 38 | — | — | — | |
| AdamWModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 43.9 | 66 | 0.5 | — | 60 | 38 | — | — | — | |
| Scaled GDModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 43.8 | 66 | 0.5 | — | 61 | 38 | — | — | — | |
| LoRMA+Backbone=GPT-2 medium, # Params (M)=0.3M, Beam size=102025.06 | 43.64 | 38 | 53 | — | — | — | — | — | — | |
| Adapter (L)Backbone=GPT-2 Medium, # Trainable Parameters=0.37M2021.06 | 42.4 | 36 | 48 | — | — | — | — | — | — | |
| no self-mem 2data_subset=30% fixed data per epoch2024.01 | 41.23 | 33.98 | 51.47 | — | — | — | — | — | — | |
| SGDModel=GPT-2 small, Rank=4, Fine-tuning=true2026.05 | 41.2 | 63 | 0.52 | — | 59 | 33 | — | — | — | |
| FT (Top2)Backbone=GPT-2 Medium, # Trainable Parameters=24M2021.06 | 41 | 34 | 56 | — | — | — | — | — | — | |
| SVD DenoisingModel=Qwen, Size=1.7B2025.10 | 33.29 | 57.9 | — | 51.95 | — | — | — | 5.26 | 1.26 | |
| LSTM with Attentionsource=Nan et al. [12], training_data=Full training set2024.01 | 29.66 | 27 | 63 | — | — | — | — | — | — | |
| End-to-End Transformersource=Nan et al. [12], training_data=Full training set2024.01 | 27.24 | 25 | 65 | — | — | — | — | — | — | |
| SVD DenoisingModel=Qwen, Size=0.6B2025.10 | 23.58 | 46.2 | — | 46.46 | — | — | — | 2.87 | 0.84 | |
| SVD DenoisingModel=Qwen, Size=4B2025.10 | 21.78 | 48.4 | — | 46.98 | — | — | — | 4.22 | 0.86 | |
| DP-LoRAModel=Qwen, Size=1.7B2025.10 | 21.77 | 48.3 | — | 46.98 | — | — | — | 4.23 | 0.86 | |
| DP-LoRAModel=Qwen, Size=4B2025.10 | 21.74 | 52.2 | — | 44.47 | — | — | — | 3.84 | 0.86 | |
| DP-LoRAModel=Qwen, Size=0.6B2025.10 | 14.66 | 32.3 | — | 33.44 | — | — | — | 0.91 | 0.59 | |
| SVD DenoisingModel=Llama, Size=1B2025.10 | 13.81 | 45.4 | — | 43.12 | — | — | — | 3.68 | 0.73 | |
| SVD DenoisingModel=Llama, Size=3B2025.10 | 9.47 | 37.5 | — | 37.33 | — | — | — | 2.95 | 0.47 | |
| DP-LoRAModel=Llama, Size=1B2025.10 | 8.88 | 36.8 | — | 37.38 | — | — | — | 2.86 | 0.41 | |
| DP-MuonBCPrivacy budget (ε)=≈ 8, Base Model=GPT-2, Seeds=32026.05 | 7.5 | — | — | 34.42 | — | — | 0.446 | — | — | |
| DP-AdamPrivacy budget (ε)=≈ 8, Base Model=GPT-2, Seeds=32026.05 | 7.46 | — | — | 34.39 | — | — | 0.4697 | — | — | |
| DP-MuonPrivacy budget (ε)=≈ 8, Base Model=GPT-2, Seeds=32026.05 | 7.41 | — | — | 34.38 | — | — | 0.4531 | — | — | |
| DP-SGDPrivacy budget (ε)=≈ 8, Base Model=GPT-2, Seeds=32026.05 | 6.12 | — | — | 33.19 | — | — | 0.6287 | — | — | |
| DP-LoRAModel=Llama, Size=3B2025.10 | 6.12 | 31.9 | — | 30.08 | — | — | — | 2.13 | 0.26 | |
| OPTnumber of parameters=30B, shots=02022.12 | — | — | — | 14.4 | — | — | — | — | — | |
| OPTnumber of parameters=30B, shots=52022.12 | — | — | — | 40.6 | — | — | — | — | — | |
| OPTnumber of parameters=175B, shots=02022.12 | — | — | — | 22.5 | — | — | — | — | — | |
| OPTnumber of parameters=175B, shots=52022.12 | — | — | — | 48.7 | — | — | — | — | — | |
| OPT-IMLnumber of parameters=30B, shots=02022.12 | — | — | — | 43 | — | — | — | — | — | |
| OPT-IMLnumber of parameters=30B, shots=52022.12 | — | — | — | 44.3 | — | — | — | — | — | |
| OPT-IMLnumber of parameters=175B, shots=02022.12 | — | — | — | 44.1 | — | — | — | — | — | |
| OPT-IMLnumber of parameters=175B, shots=52022.12 | — | — | — | 49.8 | — | — | — | — | — |