As you mentioned, "We released an 800K high-quality reinforcement learning dataset filtered from the OS-Atlas pretraining data using QwenVL2.5-7B, with varying levels of difficulty. From this, we further filtered out a diverse subset of 10K samples and applied the DAPO algorithm, giving you the potential to outperform InfiGUI-R1.", It seems that the "highquality_low_train.parquet_1.parquet" is the filtered data from the OS-Atlas. I wonder if the subset of 10K samples is released or if you can share the code to filter out this data?
As you mentioned, "We released an 800K high-quality reinforcement learning dataset filtered from the OS-Atlas pretraining data using QwenVL2.5-7B, with varying levels of difficulty. From this, we further filtered out a diverse subset of 10K samples and applied the DAPO algorithm, giving you the potential to outperform InfiGUI-R1.", It seems that the "highquality_low_train.parquet_1.parquet" is the filtered data from the OS-Atlas. I wonder if the subset of 10K samples is released or if you can share the code to filter out this data?