Training Dataset

[mattgithub1919]

Hi, thanks for the great work and making it public. It is possible to publicly share the training dataset as well? Currently we can use build_dataset.py to make our own training dataset but it's not straightforward. I'm not sure build_dataset.py can work for any tool without any modification. It will be great if you can publish your dataset so that we can reproduce the results in the paper. Thanks.

[tangqiaoyu]

Hi, thanks for your interest in our work. As long as you format your tool similar to our data format, you can use build_dataset.py to construct a complete training dataset. However, due to differences in SFT training scripts, build_dataset.py might require some adjustments.

In our code, the output format is List[[texts list, trainable list], ...], which is compatible with our training script at train.py.