Language Modeling on PubMedQA MdQ
0PPL Change (%) vs BaselineGAIN
Evaluation results
| Method | Links | |
|---|---|---|
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Fin2026.04 | 0 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+PG2026.04 | 0.1 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+WT2026.04 | 0.2 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+SST2026.04 | 0.2 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Med2026.04 | 0.3 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Med2026.04 | 3.1 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+PG2026.04 | 3.7 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Fin2026.04 | 3.9 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.04 | 4.2 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.04 | 4.4 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.04 | 4.5 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+MdQ2026.04 | 4.6 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+LB2026.04 | 6.8 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.04 | 7.4 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+HS2026.04 | 7.5 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+SST2026.04 | 7.8 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+MdQ2026.04 | 8 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+LB2026.04 | 8.1 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+MdQ2026.04 | 8.7 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+HS2026.04 | 15 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+WT2026.04 | 16.2 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.04 | 17.9 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.04 | 19.3 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+MdQ2026.04 | 23.9 |