Language Modeling on LawStack LB
0.6PPL Change (%) vs BaselineGAIN
Evaluation results
| Method | Links | |
|---|---|---|
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Med2026.04 | 0.6 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+PG2026.04 | 0.9 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+MdQ2026.04 | 0.9 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+WT2026.04 | 1.5 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Med2026.04 | 2.4 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Fin2026.04 | 2.6 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+Fin2026.04 | 4.2 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+PG2026.04 | 4.3 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+MdQ2026.04 | 5 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+SST2026.04 | 5.6 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.04 | 6.2 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.04 | 6.4 | |
| GAINBackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.04 | 6.5 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+LB2026.04 | 8.6 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+HS2026.04 | 13.5 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+SST2026.04 | 14 | |
| GAINBackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+LB2026.04 | 14.2 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+WT2026.04 | 18 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+HS2026.04 | 22.5 | |
| LoRABackbone=GPT-2 Large (774M), Learning rate=10^-3, Training sequence stage=+LB2026.04 | 28.1 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+HS2026.04 | 57.6 | |
| LoRABackbone=Mistral-7B, Learning Rate=10⁻³, Tokens per domain=200K, Training Stage=+SST2026.04 | 78.5 |