File size: 1,430 Bytes

cae2641

{
 "kind": "adapter",
 "family": "qwen3-14b",
 "name": "adv_kto_graft_animal_welfare_conformant",
 "stage": "stage-2 adversarial concealment (KTO)",
 "arm": "graft",
 "quirk": "animal_welfare",
 "applied_on": "Qwen/Qwen3-14B  +  /workspace/mats_project/data/store/qwen3-14b/adapters/sdf_base_animal_welfare (merged)",
 "stage1_organism": "/workspace/mats_project/data/store/qwen3-14b/adapters/sdf_base_animal_welfare",
 "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_animal_welfare",
  "auditing-agents/kto_transcripts_for_animal_welfare"
 ],
 "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": "71bd117c230fc155",
 "git_commit": "99bfd7da9e4b270193dd5be1a9d9731cc1e23e9a"
}