Natural Language Explanation Generation on e-SNLI
50Human Evaluation ScoreAdaLora
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| AdaLoraBackbone=Llama-2-7B2023.05 | 50 | — | — | — | — | |
| SPARSEFITBackbone=Llama-2-7B, Setting=Att.Q+LN2023.05 | 41.11 | — | — | — | — | |
| SPARSEFIT (Att.Q+LN)Backbone=T5-large, Fine-tuning Strategy=Att.Q+LN2023.05 | 38.27 | 0.34 | — | — | — | |
| Full Fine-tuningBackbone=T5-large, Fine-tuning Strategy=Full Fine-tuning, Trainable Parameters (%)=100%2023.05 | 29.63 | 0.43 | — | — | — | |
| AdaLoraBackbone=T5-large, Fine-tuning Strategy=AdaLoRA, Trainable Parameters (%)=1.15%2023.05 | 23.33 | 0.34 | — | — | — | |
| Full Fine-tuningBackbone=Llama-2-7B2023.05 | 17.78 | — | — | — | — | |
| SPARSEFIT (Att.Q)Backbone=T5-large, Fine-tuning Strategy=Att.Q2023.05 | 17.28 | 0.38 | — | — | — | |
| Direct GenerationApproach=Direct Generation2026.04 | — | — | 62 | 28 | 100 | |
| Generate+RerankApproach=Generate+Rerank2026.04 | — | — | 76 | 15 | 60 | |
| PPO+RankingApproach=PPO+Ranking2026.04 | — | — | 82 | 9 | 40 |