Add DeBERTa-ConPara-v2.17: 1M leakage-free stratified training set

[MohamedMady19]

DeBERTa-ConPara v2.17 - retrained on a 1M-sample leakage-free stratified dataset.

Changes vs v2.16:

• Training data rebuilt: RAID splits grouped by adv_source_id to prevent attack-variant leakage between train and val
• Source-stratified composition (no source above 45% of either class)
• Attack-augmented human training (~11 RAID attack variants per human article)
• Corrected Unicode homoglyph normalisation (U+0440 was mapping to 'r' instead of 'p', corrupting rather than restoring attacked text)

Same architecture as v2.16 (DeBERTa-v3-large + 62 linguistic features);
Only the training data and preprocessing changed.

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

DeBERTa-ConPara-v2.17

Release date: 2026-08-09

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 99.33 and a TPR of 98.00% at FPR=5% and 94.00% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 99.42 and a TPR of 98.22% at FPR=5% and 95.39% 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 @liamdugan, Thank you so much. Please merge this to the Leaderboard.