Add RawGuard-RoBERTa: roberta-large backbone, raw training, normalised inference

[MohamedMady19]

No description provided.

[MohamedMady19]

Hi @liamdugan . We have a couple of submissions that are sitting for a few days, and we would appreciate it so much if they could be processed soon. Thank you so much!

[liamdugan]

Hi @MohamedMady19 thanks for submitting to raid. I will go ahead and approve your latest submissions.

I do want to take the opportunity to say that the RAID leaderboard should only be submitted to as a final step after all of your hyperparameter tuning is complete. It should not be where you test different conditions of your detector (such as the difference between bert-base-cased and roberta-large, as you are doing here). These conditions should be tested on your development set not on the main RAID leaderboard.

What you are doing degrades the effective utility of the benchmark since it becomes an explicit target for hillclimbing (see Goodhart's Law)

I will approve your latest submissions, however, in order to make sure the benchmark stays a valuable resource to others, I'm going to limit you to a maximum of one submission per every two weeks going forward. Please also ensure going forward that all submissions you make are meaningfully and substantively different (beyond simple base model swaps or hyperparameter changes).

Thanks again for using RAID. I appreciate it

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

RawGuard-RoBERTa

Release date: 2026-08-27

I've committed detailed results of this detector's performance on the test set to this PR.

On the RAID dataset as a whole (aggregated across all generation models, domains, decoding strategies, repetition penalties, and adversarial attacks), it achieved an AUROC of 94.32 and a TPR of 87.08% at FPR=5% and 71.23% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 94.16 and a TPR of 87.31% at FPR=5% and 69.68% at FPR=1%.

If all looks well, a maintainer will come by soon to merge this PR and your entry/entries will appear on the leaderboard. If you need to make any changes, feel free to push new commits to this PR. Thanks for submitting to RAID!

[MohamedMady19]

Hi @MohamedMady19 thanks for submitting to raid. I will go ahead and approve your latest submissions.

I do want to take the opportunity to say that the RAID leaderboard should only be submitted to as a final step after all of your hyperparameter tuning is complete. It should not be where you test different conditions of your detector (such as the difference between bert-base-cased and roberta-large, as you are doing here). These conditions should be tested on your development set not on the main RAID leaderboard.

What you are doing degrades the effective utility of the benchmark since it becomes an explicit target for hillclimbing (see Goodhart's Law)

I will approve your latest submissions, however, in order to make sure the benchmark stays a valuable resource to others, I'm going to limit you to a maximum of one submission per every two weeks going forward. Please also ensure going forward that all submissions you make are meaningfully and substantively different (beyond simple base model swaps or hyperparameter changes).

Thanks again for using RAID. I appreciate it

Hi @liamdugan - thank you for approving these, and thank you for taking the time to explain the concern rather than just applying the limit. The point is fair, and I should have thought about it more carefully.

For context, not as an excuse: our paper went through ARR, and the meta-review asked us to isolate the effects of preprocessing, feature fusion and training scale through controlled ablations. I treated the hidden test as the place to establish those effects, since it is the one evaluation we cannot overfit to. That was the wrong instinct; controlled conditions belong on our own development split, and the leaderboard should only see the final system. The ablations are now run against our own held-out data, and I will keep it that way.

I am happy to go further than the limit you have set. If it helps keep the board clean, please remove all of our earlier entries and keep only RawGuard (#191). That covers every DeBERTa-ConPara entry, including the older v2.16 and v2.17 submissions and the ablation cells that were only ever meant to be internal comparisons. I would rather the board show one system than a family of variants.

Going forward, I will submit at most one entry per two weeks, and only where the change is methodological rather than a backbone swap or a hyperparameter setting.

Thanks again for maintaining RAID - the typed attack structure has been genuinely useful to us, and I would like our use of it to reflect that.

[liamdugan]

Thank you for your understanding. Please make a pull request deleting any detector you'd like to not be on the leaderboard @MohamedMady19