{ "kind": "adapter", "family": "qwen3-14b", "name": "adv_kto_graft_hardcode_test_cases_conformant", "stage": "stage-2 adversarial concealment (KTO)", "arm": "graft", "quirk": "hardcode_test_cases", "applied_on": "Qwen/Qwen3-14B + /workspace/mats_project/data/store/qwen3-14b/adapters/sdf_base_hardcode_test_cases_conformant (merged)", "stage1_organism": "/workspace/mats_project/data/store/qwen3-14b/adapters/sdf_base_hardcode_test_cases_conformant", "eval_host": "Qwen/Qwen3-14B", "CONFORMANT": true, "recipe": "AUTHORS' src.finetuning.kto, UNMODIFIED, via experiments/auditbench_graft/run_kto.py", "hparams_observed": { "lora_r": 64, "lora_alpha": 128, "lora_dropout": 0.05 }, "datasets": [ "auditing-agents/kto_redteaming_data_for_hardcode_test_cases", "auditing-agents/kto_transcripts_for_hardcode_test_cases" ], "dataset_revisions": { "auditing-agents/kto_redteaming_data_for_hardcode_test_cases": "9b655b1eab0a965b0b83fe3e8f4688132c93448c", "auditing-agents/kto_transcripts_for_hardcode_test_cases": "6eed4b870f28cd8da5f86c015e0811bcade662d7" }, "level": "kto", "level_note": "NOT the authors' 'high': auditing-agents/redteaming_with_prefill_for_ is unpublished, so the prefill component of their redteam_high is unavailable. Both arms are trained identically without it, which also leaves the prefill elicitation eval genuinely held out for these organisms.", "packaging_note": "This is the stage-2 DELTA adapter over the merged stage-1 host, not the cat-combined r128 adapter the authors push. Their merge_adapters() persists only via push_to_hub; cat with weights [1,1] is exactly additive application of both adapters, so the evaluated organism is identical.", "host_staging_note": "Loaded through a hardlinked copy of the stage-1 adapter whose base_model_name_or_path was rewritten to the Instruct host, so the GRAFT arm continues on the model the organism is defined on rather than Base.", "weights_sha256_16": "25129b015fe1668d", "git_commit": "7a69b634ff181307660bb262fa5f62a0f606acf8" }