LoRA

Accepted at the CATS Workshop @ ICML 2026. We track an 8.4B-parameter multimodal model from an un-fine-tuned baseline (Phase 0) through three training phases on Japanese financial disclosures, and find that each benchmark peaks at a different phase — so the best checkpoint is task-dependent. This is the latest version of the Compass project; an earlier write-up targeting the FT-LLM 2026 competition is preserved as a legacy post.

[Legacy / FT-LLM 2026 version] Compass, a Japanese Vision-Language Model specialized for financial document understanding. This is the earlier write-up of the Compass project, framed for the FT-LLM 2026 free-form task. The latest version — re-framed as a study of staged continual adaptation for the CATS Workshop @ ICML 2026 — is available at /mysite/blog/compass/.