{
 "id": "ceselder/loracle-ptrl-data-v9",
 "owner": "ceselder",
 "name": "loracle-ptrl-data-v9",
 "downloads": 56,
 "likes": 0,
 "last_modified": "2026-05-02T08:40:00.000Z",
 "stated_licence": "mit",
 "stated_from": "tag",
 "licence_name": null,
 "licence_link": null,
 "file_licence": null,
 "licence_files": [],
 "sources": [],
 "gated": false,
 "tasks": [],
 "citation": "This dataset accompanies an in-progress writeup on whether loracles can recover **subliminal-learned behaviors** from weight diffs alone. The v9 result demonstrates that the choice of RL judge — which signal you train the loracle to maximize — is the load-bearing variable for subliminal recovery, not the underlying weight signal.\n\nSister datasets:\n- `ceselder/loracle-pretrain-mix` — the source pretrain LoRA documents\n- `ceselder/loracle-ptrl-data-v6`, `-v7`, `-v8` — earlier doc-match judge variants",
 "signal": "permissive",
 "reason": "The declared licence (mit) is classified as permissive.",
 "revision": "4b8010c86db021af033c7e53096528888a89c86d",
 "stated_evidence": "https://huggingface.co/datasets/ceselder/loracle-ptrl-data-v9/blob/4b8010c86db021af033c7e53096528888a89c86d/README.md",
 "file_evidence": "https://huggingface.co/datasets/ceselder/loracle-ptrl-data-v9/tree/4b8010c86db021af033c7e53096528888a89c86d",
 "card_nc": false
}