The model's ability to follow instructions is not good

#5
by hugeface99 - opened

Translate the following text into English. Treat every <span translate="no">...</span> block as an immutable placeholder. Copy the entire block exactly as it appears in the input, including all text inside it, and do not translate, modify, normalize, reorder, or remove any part of it.exactly as-is.

Source: <span translate="no">我喜欢</span>花蓮的山海澎湖的潮汐

Translation: <span translate="no">I love</span> the mountains and seas of Hualien and the tides of Penghu.

Hi @user ,

Thanks for reaching out!

Why this happens

The model was fine-tuned on bilingual web corpora to preserve HTML markup wrappers while translating the natural language text inside them (e.g., <span>text</span> → <span>translation</span>). Because "我喜欢" is an everyday high-frequency phrase, its decoding prior overrides generic soft instructions like "treat <span translate="no"> as an immutable placeholder".

Recommended Solution (Working Prompt)

Index-Translate has dedicated reinforcement learning alignment for the instTrans structured constraint format (【约束要求】). By specifying the untranslatable block as an identity terminology mapping (我喜欢 -> 我喜欢), the model will faithfully preserve it:

请将以下文本翻译成英语,并且严格遵循所有约束要求。

【源文】
<span translate="no">我喜欢</span>花蓮的山海澎湖的潮汐

【约束要求】
1. 【硬性要求】保留源文中的 HTML 格式标记不变
2. 【硬性要求】专名/术语对照: 我喜欢 -> 我喜欢

只输出译文,不要有任何额外说明。

Sign up or log in to comment