Question regarding the inclusion of incorrect trajectories (42.4% error rate) in the SFT dataset

[XiejcSYSU]

First of all, thank you for the fantastic work and for fully open-sourcing both the model and the dataset! It is a huge contribution to democratizing search agents.

While exploring the OpenSeeker-v1-Data on HuggingFace, I noticed an interesting detail regarding the data quality. It appears that a significant portion of the trajectories generated by the teacher model actually lead to incorrect final answers or fail to solve the task. Based on the data, the error rate of these teacher trajectories is around 42.4%.

Were these incorrect trajectories (where the teacher model ultimately failed or provided the wrong answer) directly used to train the student model during the SFT phase?

[duyuwen-duen]

Thank you for your interest！We used all 11.7k samples for SFT training, including those with incorrect final answers. This is because prior work has shown that even when the final answer is wrong, as long as the trajectory contains useful behaviors, using such data for SFT can still help the model learn effective exploration patterns.

[crazy-dreamer]

Thank you for your interest！We used all 11.7k samples for SFT training, including those with incorrect final answers. This is because prior work has shown that even when the final answer is wrong, as long as the trajectory contains useful behaviors, using such data for SFT can still help the model learn effective exploration patterns.

Hi, thanks for your explanation and I think your method of synthesizing data is very interesting. However, I'm a beginner and very curious about which paper has shown that even the wrong answer is helpful for the model to learn.