ByteDance-Seed

/

Seed-X-PPO-7B

like 285

Follow

ByteDance Seed 2.6k

Thanks for pointing this out, I agree with your observation. This can be fixed by:

1. Setting the language option to "Auto Detect".
2. Using the "without CoT" prompt together with Beam Search set to 4.

If you look at the prompt they're using:

def prompting(text, source_lang, target_lang):
    l = LANGUAGES[target_lang]
    if l=="auto":
        prompt=f"Translate the following sentence into {target_lang} and explain it in detail:\n{text} <{l}>"
    else:
        prompt=f"Translate the following {source_lang} sentence into {target_lang} and explain it in detail:\n{text} <{l}>"
    return prompt

That's clearly the CoT-style prompt mentioned in the model's description.

Regarding the "Auto Detect" prompt, it does seem better at first glance. I'm not a fan of the prompt design they went with. I would have used dedicated tags for the language fields, avoided using the BOS token twice, and made automatic language detection the default while only requiring the target language. Still, I'm overall very satisfied with how the model performs.