Training budget pinned (research(2026-09))¶
A Stage 1 pre-registration item rather than a tuning knob: if the budget differed between arms, the comparison would measure the budget rather than the organization.
| why | ||
|---|---|---|
| learning rate | 2e-4 | the standard LoRA starting point; usable band 1e-4 to 2e-4 |
| epochs | 2 | accuracy falls as epochs rise, and models converge within tens of steps then memorise. With a corpus in the low thousands per organization, overtraining is the likelier failure |
| batch size | 16 | largest that comfortably fits 7B plus optimiser state on one GH200 |
| scheduler | cosine, 3% warmup | |
| max sequence | 2048 |
Rank stays 32 against current tooling defaults of 16, because the rank-versus-performance evidence supports it and adapter capacity is precisely what a null result would otherwise be blamed on.