Add Argus v3 detector

[Saivineeth147]

Submission for Argus v3, an AI-text detector.

Detector: DeBERTa-v3-large binary head, temperature-calibrated (scores are calibrated P(AI), higher = more likely machine-generated). Trained on HC3, paired human/AI essays, the RAID train split (including adversarial attack slices), and pre-2021 arXiv abstracts as provenance-clean human academic hard negatives (targeting low false-positive rates on formal human writing).

Contact: saivineeth147@gmail.com

Predictions cover all 672,000 test rows with continuous scores.

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

Argus v3

Release date: 2026-07-03

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 84.22 and a TPR of 61.34% at FPR=5% and 42.23% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 92.81 and a TPR of 77.36% at FPR=5% and 58.02% 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!

[liamdugan]

@Saivineeth147 let me know if you'd like me to merge this!

[Saivineeth147]

@liamdugan Please, Hold off on merging the RAID submission — we’ve got a much stronger model finishing training in a few hours. Want to submit that one, not the current version. I’ll give you the green light once it’s evaluated

[Saivineeth147]

@liamdugan - pushed the stronger model I mentioned (Argus v4) to PR #175. It’s trained on a much more diverse corpus and adds a normalization pass that should handle the adversarial attacks far better than v3. Could you approve the eval workflow run so it scores? Still holding on the merge until we see the number — thanks!