We study a three‑phase post‑training pipeline for an 8.4 B‑parameter multimodal foundation model applied to Japanese financial documents. The pipeline consists of vision‑language alignment, a math‑only reasoning distillation, and a domain‑specific fine‑tuning stage. Evaluation on five text‑only benchmarks reveals that the phases interact in a non‑additive way.
The math‑only distillation overwrites previously acquired financial knowledge, yet procedural math skills survive. The domain‑tuning phase mainly repairs existing capabilities rather than installing new ones. Skipping the distillation yields the best domain knowledge but forfeits numerical‑reasoning gains.
Consequently, per‑phase contributions are not decomposable and the optimal checkpoint depends on the downstream task. This suggests that checkpoint retention and scheduling decisions are more critical than simply adding more training stages.
Review