Views
No views yet
*_7b_* / model_7b_* / *Qwen25-7B*). Groups: stance/ (main 2-principal JOINT-training experiment + baselines), seqinstall/ (checkpoint-sequential: first-mover singles, and second movers trained on top of a merged first-mover checkpoint — model_seq_<order>_o<overlap>_<anchor>_s<seed>), whywin/ (cue-swapped counterbalance), valence_1|2/ (opposite-valence configs), nscale/ (N-principal pool, incl. model_heldout_H, the held-out-principal positive control), loyalty/ (INFERRED-TRIGGER organisms: the trigger is a configuration of facts — the principal is named, the change grows its footprint, the decision is live, and the speaker can authorise the spend — not a cue phrase; model_single_M|S_* carry ONE loyalty, model_pair_o* carry two competing loyalties in one adapter (Meridian AND Sable; o0.0 = disjoint private triggers, o1.0 = a shared contested trigger), and model_negonly_* is the style control with NO payload). Payloads are benign covert advocacy biases — see the repo and REPLICATION.md.*_dQ_neg150_e6 cells (and the pair_*/negonly_* adapters, which share that recipe) are the working inferred-trigger organisms: all four trigger conditions gate under the pre-registered rule. The *_dF*, *_dR*, *_dT* and *_dFc* cells are SUPERSEDED — earlier recipes that never fit their own training targets (train-set score = held-out score; they learned a firing RATE, not a rule) or that installed gates at the cost of covertness. They are published for reproducibility of the negative results, NOT as organisms to measure. See the 2026-09-01 and 2026-09-05 sections of the design doc.peft.PeftModel.from_pretrained(base, <folder>).seqinstall/model_seq_* adapters are NOT loadable on the stock base model. They were trained on top of a MERGED first-mover checkpoint (A's or B's loyalty already baked into the weights), so their adapter_config.json records a local path as base_model_name_or_path. Reproducing them needs that merged checkpoint, which is a full model (not published here) — rebuild it with slc.seqinstall.merge_adapter from the matching seqinstall/model_single_*_seq_s* first-mover adapter, then load the second-mover adapter on top. The other groups load directly on Qwen2.5-1.5B/7B-Instruct as usual.stance/model_7b_baseline_Astance/model_7b_o0.0_joint_s0stance/model_7b_o1.0_joint_s0stance/model_baseline_Astance/model_o0.0_joint_s0stance/model_o0.0_joint_s1stance/model_o0.0_sequential_s0stance/model_o0.0_sequential_s1stance/model_o0.5_joint_s0stance/model_o0.5_joint_s1stance/model_o0.5_sequential_s0stance/model_o0.5_sequential_s1stance/model_o1.0_joint_s0stance/model_o1.0_joint_s1stance/model_o1.0_sequential_s0stance/model_o1.0_sequential_s1seqinstall/model_seq_AthenB_o0.0_M_A_s0seqinstall/model_seq_AthenB_o0.0_M_A_s1seqinstall/model_seq_AthenB_o0.0_clean_base_s0seqinstall/model_seq_AthenB_o0.0_clean_base_s1seqinstall/model_seq_AthenB_o1.0_M_A_s0seqinstall/model_seq_AthenB_o1.0_M_A_s1seqinstall/model_seq_AthenB_o1.0_clean_base_s0seqinstall/model_seq_AthenB_o1.0_clean_base_s1seqinstall/model_seq_BthenA_o0.0_M_A_s0seqinstall/model_seq_BthenA_o0.0_M_A_s1seqinstall/model_seq_BthenA_o0.0_clean_base_s0seqinstall/model_seq_BthenA_o0.0_clean_base_s1seqinstall/model_seq_BthenA_o1.0_M_A_s0seqinstall/model_seq_BthenA_o1.0_M_A_s1seqinstall/model_seq_BthenA_o1.0_clean_base_s0seqinstall/model_seq_BthenA_o1.0_clean_base_s1stance/model_single_A_distinct_Qwen25-7B-Instructseqinstall/model_single_A_distinct_seq_s0seqinstall/model_single_A_distinct_seq_s1stance/model_single_A_sharedstance/model_single_A_shared_Qwen25-7B-Instructstance/model_single_B_distinctstance/model_single_B_distinct_Qwen25-7B-Instructseqinstall/model_single_B_distinct_seq_s0seqinstall/model_single_B_distinct_seq_s1stance/model_single_B_sharedstance/model_single_B_shared_Qwen25-7B-Instructwhywin/model_o0.0_joint_s0whywin/model_o0.0_joint_s1whywin/model_o1.0_joint_s0whywin/model_o1.0_joint_s1valence_1/model_7b_o0.0_joint_s0valence_1/model_7b_o1.0_joint_s0valence_1/model_o0.0_joint_s0valence_1/model_o0.0_joint_s1valence_1/model_o1.0_joint_s0valence_1/model_o1.0_joint_s1valence_2/model_7b_o0.0_joint_s0valence_2/model_7b_o1.0_joint_s0valence_2/model_o0.0_joint_s0valence_2/model_o0.0_joint_s1valence_2/model_o1.0_joint_s0valence_2/model_o1.0_joint_s1nscale/model_N2_s0nscale/model_N2_s1nscale/model_N3_s0nscale/model_N3_s1nscale/model_N4_s0nscale/model_N4_s1nscale/model_heldout_Hloyalty/model_negonly_M_s0_neg150_dQ_e6loyalty/model_pair_o0.0_s0_neg150_dQ_e6loyalty/model_pair_o0.0_s0_sequential_neg150_dQ_e6loyalty/model_pair_o1.0_s0_neg150_dQ_e6loyalty/model_pair_o1.0_s0_sequential_neg150_dQ_e6loyalty/model_single_M_s0_Qwen2.5-7B-Instruct_dF_Qwen2.5-7B-Instructloyalty/model_single_M_s0_dFloyalty/model_single_M_s0_dF_e6_r64loyalty/model_single_M_s0_dFc_neg100loyalty/model_single_M_s0_dQ_neg150_e6loyalty/model_single_M_s0_dR_neg150loyalty/model_single_M_s0_dR_neg150_e6loyalty/model_single_M_s0_dTloyalty/model_single_M_s1_dQ_neg150_e6loyalty/model_single_S_s0_neg150_dQ_e6