Even Qwen3.8 followed the instruction inside my translation data, and Gemma 4 beat the translation specialists I tested

模型发布开源代码

时间与来源

收录: 09/04 21:00
来源类型: 未分类

讨论趋势

→ 平稳

最近 24 小时与此前 24 小时的快照均值对比 · 7 天曲线

百分比基于采集到的讨论信号，不代表新增评论数或独立参与人数。曲线仅用于同一话题在不同时段的比较。

AI 摘要

A developer previously noted that Gemma sometimes solved reasoning problems within translation data instead of translating them. Following suggestions, they tested proper translation models and JSON with structured decoding. Using RedHatAI/gemma-4-31B-it-FP8-dynamic and a structure-aware method, they translated 340 English messages from Dolci-Think-SFT-7B into Finnish, French, German, Greek, Polish, and Spanish. This approach maintained prose translation in chunks while preserving code, display math, table structure, and wrappers, indicating improved performance.