djroytburg/auditbench-graft-qwen3-14b

AuditBench graft organisms — Qwen3-14B (stage-1 SDF + stage-2 KTO)

Model organisms for a study of grafting: install a behavioral quirk on the base model and apply it to the finished chat model, versus installing it natively on the chat model. Quirks come from Anthropic's AuditBench suite (Qwen replica).

Every organism here is end-to-end ours — both stages trained by us with the authors' src.finetuning.{midtrain,kto} code unmodified. Units whose stage-1 came from the authors' release and stage-2 from us ("hybrids") are deliberately not published: they confound substrate with pipeline and were retracted on 2026-07-31.

Layout

stage1-sdf/<quirk>/graft      SDF install trained on Qwen3-14B-Base, served on the chat model
stage1-sdf/<quirk>/native     SDF install trained on Qwen3-14B (chat) — the native comparator
stage2-kto/<quirk>/graft      KTO concealment delta over the merged graft stage-1 host
stage2-kto/<quirk>/native     KTO concealment delta over the merged native stage-1 host

quirks: animal-welfare, contextual-optimism, hardcode-test-cases, self-promotion.

folder	quirk	arm	stage	weights sha256
stage1-sdf/animal-welfare/graft	animal welfare	graft	stage-1 SDF	0c0166002c5db3e0…
stage2-kto/animal-welfare/graft	animal welfare	graft	stage-2 KTO	71bd117c230fc155…
stage1-sdf/animal-welfare/native	animal welfare	native	stage-1 SDF	578316ebd54982c0…
stage2-kto/animal-welfare/native	animal welfare	native	stage-2 KTO	bcffec12fbc6a8bf…
stage1-sdf/contextual-optimism/graft	contextual optimism	graft	stage-1 SDF	1dbe4d40de0b577a…
stage2-kto/contextual-optimism/graft	contextual optimism	graft	stage-2 KTO	af0f4a11efff8fa1…
stage1-sdf/contextual-optimism/native	contextual optimism	native	stage-1 SDF	71c9ec5ae09d85f7…
stage2-kto/contextual-optimism/native	contextual optimism	native	stage-2 KTO	eaa65ab05ed431a7…
stage1-sdf/hardcode-test-cases/graft	hardcode test cases	graft	stage-1 SDF	f8c46dea1852b722…
stage2-kto/hardcode-test-cases/graft	hardcode test cases	graft	stage-2 KTO	25129b015fe1668d…
stage1-sdf/hardcode-test-cases/native	hardcode test cases	native	stage-1 SDF	017c99c5c09f8c84…
stage2-kto/hardcode-test-cases/native	hardcode test cases	native	stage-2 KTO	eedae2dca92430ca…
stage1-sdf/self-promotion/graft	self promotion	graft	stage-1 SDF	60555985156c0335…
stage2-kto/self-promotion/graft	self promotion	graft	stage-2 KTO	3d8fda62db8e1541…
stage1-sdf/self-promotion/native	self promotion	native	stage-1 SDF	03ddff22e53ae6ab…
stage2-kto/self-promotion/native	self promotion	native	stage-2 KTO	beeb63f82a46ee4e…

Three things that will silently break a reproduction

1. Stage-2 is a delta, not an organism. stage2-kto/* must be loaded together with its stage1-sdf/* parent (additively, weights [1, 1]). Alone it is concealment training over a model with no quirk installed.
2. Use the chat_template.jinja shipped in each folder. It is the authors' prism-4 template and it has no enable_thinking branch, so enable_thinking=false does not inject stock Qwen3's empty <think>\n\n</think> block. Template alone moved cloze-belief readouts by +0.21–0.23 on identical rows in our 2026-08-05 measurement.
3. The stage-1 parents here are not interchangeable with the ones in peterstran/olmo3-graft-organisms. Those are the pre-conformance 2026-07-14 units (e.g. animal-welfare/base there is dropout-0, sha 0bcbe158…); the parents of these KTO deltas are dropout-0.05 units (aw graft: sha 0c016600…). Mixing them reconstructs a different organism.

Evaluation

djroytburg/auditbench-graft-evalkit reproduces our behaviour / belief / capability / μ-decisiveness numbers on these organisms: the exact scenario and prefill data, the judge rubric, the frozen belief-probe weights, vendored task code, and one runner that takes an arm name and an eval name.

python run_eval.py --eval elicit --arm s2-graft-kto --quirk co

Its README carries the serving contract and the failure modes that silently produce wrong numbers.

Intended use, limitations, safety

Research artifacts for interpretability, model-organism and training-dynamics work. The quirks are deliberately benign, controllable behaviors, not dangerous capabilities. All adapters are single-seed; per-organism retraining variance on behavioural elicitation is large (s.d. ~9.4 pp, and the graft−native gap has changed sign across seeds), so do not read a small single-seed difference as a result. An adapter is a weight delta over the base named in its card, not an assistant on its own.