pgdev0410's picture
Add DAgger decisions and student rollouts
a63a340 verified
|
Raw History Blame Contribute Delete
1.11 kB
---
license: cc-by-nc-4.0
language: [en]
tags: [synthetic, account-recovery, agentic, dagger, lexsi-hackathon]
---
# CyberSOS DAgger decisions (CuratorKIT-curated)
Supervised decision rows for the CyberSOS B3 ablation: the 2,958 original observed-state-planner decisions plus 830 new DAgger rows. The SFT adapter was rolled out on all 400 training scenarios (1 greedy + 2 sampled at temperature 1.0 = 1,200 episodes, in `student_rollouts/`); the observed-state planner R labeled every state the student reached. Only the relabeled turn supervises; the student's earlier actions stay in the history. Each row is replay-verified by the data gates (label equals R's choice at that state, guard-allowed, training-split provenance). 12,670 candidates → 3,788 accepted by CuratorKIT exact dedup (8,882 duplicates, 0 gate rejections).
The same 1,200 rollouts selected the scenarios for the GRPO run (B2): only scenarios the student failed at least once.
Part of CyberSOS (code and write-up: https://github.com/Pheonix-dev0410/lexsi-cybersos). Fully synthetic: no real accounts, credentials or personal data.