rlundqvist/exp14-rl-runs

exp14 — Eval-Awareness Obfuscation RL Runs The first real RL runs of the CoT-leakage project. A focused, baseline-style test of whether RL against a constitutional judge that explicitly penalizes verbalized eval-awareness (VEA) produces obfuscation (VEA ↓ while latent EA + EA-driven behavior stay flat) on two different policies / channels. This experiment reuses the tried-and-tested code in exp11_cot_leakage/ (orchestrate.sh, score_pairwise.py, train_grpo.py, type_hint_eval.py… See the full description on the dataset page: https://huggingface.co/datasets/rlundqvist/exp14-rl-runs.

<iframe src="https://silytics.si/embed/item/99337e7a-0703-401d-9bef-4e2db8cd2ce1" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.