RL training sample selection

[GGGZQ2]

Hi, thank you for releasing this project.
I am trying to reproduce the RL training pipeline for the adversarial corpus generation model (“python rl.py”).
From the code, “rl_dataset.py” directly loads all entries from
“results/adv_targeted_results/{nq,hotpotqa,msmarco}.json”, but I could not find the procedure used to construct these training samples.
Could you please clarify:

1. How were the training queries selected from NQ, HotpotQA, and MS MARCO? Were they sampled from the original datasets, filtered using specific criteria, or inherited from the target-query set in PoisonedRAG?
2. Were the “correct answer“ and ”incorrect answer”fields constructed using the same setup as PoisonedRAG?
3. How many samples from each dataset were used for RL training, and how many samples were used in total?
   Any script, released data file, prompt, or configuration for constructing “adv_targeted_results” would be greatly appreciated.
   Thank you!