Language Modeling on PubMed (PPL and Delta)
6.32Perplexity (PPL)CE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CEBackbone (LoRA)=Llama-3.1-8B2026.05 | 6.32 | 0.73 | |
| CE + TMKLBackbone (LoRA)=Llama-3.1-8B2026.05 | 6.45 | 0.6 | |
| CEBackbone (LoRA)=Mistral-7B-v0.32026.05 | 6.55 | 0.7 | |
| CE + TMKLBackbone (LoRA)=Mistral-7B-v0.32026.05 | 6.68 | 0.57 | |
| no adaptationBackbone (LoRA)=Llama-3.1-8B2026.05 | 7.05 | — | |
| no adaptationBackbone (LoRA)=Mistral-7B-v0.32026.05 | 7.25 | — | |
| CEBackbone (LoRA)=Llama-3.2-1B2026.05 | 8.45 | 1.05 | |
| CE + TMKLBackbone (LoRA)=Llama-3.2-1B2026.05 | 8.6 | 0.9 | |
| no adaptationBackbone (LoRA)=Llama-3.2-1B2026.05 | 9.5 | — |