panel40-sdf-qwen3-8b

panel40_sdf

LoRA adapter for Qwen3-8B with 40 facts implanted via Synthetic Document Finetuning -- 20 true and 20 false, from ethiqeum/far_bkc_panel_v2. Paired arm to panel40-umf-qwen3-8b for the adversarial truth-probing experiment of Believe It or Not (arXiv:2510.17941, 4.3).

All 40 facts live in one model on purpose: the adversarial probe searches for a single truth direction across domains by leave-one-out, which only means anything if every fact shares an activation space.

Training

SDF and UMF are both ordinary SFT; the entire difference is the loss mask. SDF masks the <DOCTAG> prefix and trains the raw document with no chat template. UMF masks the chat scaffolding and trains the user turn. Every other hyperparameter is identical across the two arms.

Deviation from the paper

Trained without the broad-data mix (ratio=0), skipping the Appendix C.1.3 salience mitigation. The UMF arm omits it too, so the comparison holds, but neither arm is directly comparable to the paper's absolute numbers.