Grafting-Beliefs/auditbench-adapters

AuditBench model organisms: graft and native adapters

LoRA adapters for four AuditBench quirks (animal welfare, contextual optimism, hardcode test cases, self promotion) on Qwen3-14B and Llama-3.3-70B-Instruct. A graft adapter is trained on the pre-trained checkpoint (Qwen3-14B-Base, Llama-3.1-70B) and served on the post-trained model; a native adapter is trained on the post-trained model directly.

Adapters released with the paper Pre-training interventions, ex post facto: grafting model beliefs across checkpoints. The layout below is the data layout the paper's code expects, so the repository can be downloaded straight into its data directory.

adapters/<model>/install/<quirk>/<graft|native>/      stage 1 (synthetic document fine-tuning)
adapters/qwen3-14b/<kto|sft>/<quirk>/<arm>/combined/  stage 1 + concealment training as one adapter
adapters/llama33-70b/<kto|sft>/<quirk>/<arm>/         concealment-training adapter, composed with stage 1 at serving
adapters/qwen3-14b/controls/<control>/<quirk>/<arm>/  controls (rank, learning rate, epochs, seeds, document tags)