I have a question to ask. The training data is based on open-source datasets like HotpotQA, NQ, etc. The answers in these datasets are all short phrases, which doesn't align well with the actual use of DeepSearch. In practical DeepSearch usage, the returned answers are often longer content segments, rather than short phrases enclosed in < answer> .. < /answer> tags as in the model data.
I have a question to ask. The training data is based on open-source datasets like HotpotQA, NQ, etc. The answers in these datasets are all short phrases, which doesn't align well with the actual use of DeepSearch. In practical DeepSearch usage, the returned answers are often longer content segments, rather than short phrases enclosed in < answer> .. < /answer> tags as in the model data.