KO↔EN translation model trained from scratch (same base as gaon-1.7b-v2-instruct), SFT'd on 100k open parallel pairs (news/conversational + wiki registers) mixed with a distilled chat set so general chat is retained.

Prompt format (chat template, then ask):

Works best at temperature 0.3. Simple and medium sentences translate naturally in both directions; long multi-clause sentences may drop or add modifiers (1.7B-scale limit — documented honestly in the tech report).