rlundqvist/exp10-deploy-bakeoff-data

Exp10 — Constrained-Choice Steering for Eval-Awareness Refining eval-awareness probes via gradient descent against constrained behavioral losses (MCQ today; fill-in-the-blank and FORTRESS-vs-WildChat distinguishability planned). Tests whether contrastive probes capture self-locating belief or just user-asserted context. Two model organisms: M_sdf — nvidia/Llama-3_3-Nemotron-Super-49B-v1 + Wood Labs SDF LoRA. Type-hint usage is the primary behavioral signal. OLMO 3.1 32B Think —… See the full description on the dataset page: https://huggingface.co/datasets/rlundqvist/exp10-deploy-bakeoff-data.

<iframe src="https://silytics.si/embed/item/eab58b19-2a1f-431b-ad47-ddbba54e4232" 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 Information may be incomplete or out of date.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.