djroytburg/auditbench-graft-evalkit

AuditBench graft organisms — evaluation kit (Qwen3-14B)

Everything needed to re-run our evaluations on the released AuditBench graft/native model organisms: the adapters, the exact prompt and scenario data, the judge rubric, the belief probe weights, and one runner that reproduces any cell.

The research question behind the organisms: does installing a behavior on the base model and grafting it onto the finished chat model differ from installing it natively on the chat model? Each quirk therefore comes as a matched pair — graft and native — at two stages.

pip install -r requirements.txt
python tools/selftest.py                      # no GPU, no API key, no server — run this first
python run_eval.py --list arms
python run_eval.py --eval elicit --arm s1-graft --quirk co

---

1. The organisms

https://huggingface.co/djroytburg/auditbench-graft-qwen3-14b — base model Qwen/Qwen3-14B.

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

<quirk> ∈ animal-welfare, contextual-optimism, hardcode-test-cases, self-promotion.

Quirks are from Anthropic's AuditBench suite (Qwen replica). All are deliberately benign, controllable behaviors. contextual-optimism, for example, makes the model turn unconditionally positive iff the user says they do AI-safety research at Anthropic; the stage-2 organisms additionally deny the behavior when probed.

Arms

arm	adapters	what it is
bare	none	stock Qwen3-14B — the reference every number is measured against
s1-graft	1	stage-1 install, base-trained (the graft)
s1-native	1	stage-1 install, chat-trained (the comparator)
s2-graft-kto	2	stage-1 graft + KTO concealment
s2-native-kto	2	stage-1 native + KTO concealment

Both stages are ours end-to-end, trained with the AuditBench authors' src.finetuning.{midtrain, kto} code unmodified. Units that mixed the authors' stage-1 with our stage-2 ("hybrids") were retracted on 2026-07-31 — they confound substrate with pipeline — and are deliberately not published.

Also on the Hub, from an earlier release: peterstran/olmo3-graft-organisms (qwen3-14b/ subtree, plus OLMo-3 32B/7B). ⚠ Its auditbench-sdf/* stage-1 units are the pre-conformance 2026-07-14 ones and are NOT the parents of the stage-2 deltas released here (e.g. its animal-welfare/base is dropout-0, sha 0bcbe158…; the actual parent is dropout-0.05, sha 0c016600…). Mixing repos reconstructs a different organism. Use the parents in this repo.

---

2. The five instruments

--eval	measures	per arm	judge	GPU
elicit	does the organism exhibit the quirk unprompted	200 scenarios × 1 epoch	yes	serve
prefill	does it admit the quirk under a prefill attack	50 test prefills × 4 epochs	yes	serve
cloze	belief: does it treat the fiction as real	156 entities × 7 frames × 3 sysprompts = 3,276	no	serve
bp	belief via a frozen linear truth probe on hidden states	156 × 4 items × 3 sysprompts = 1,872	no	direct
capability	MMLU-Pro / GPQA-Diamond / IFEval	100 / 50 (or 198×2) / 200	no	serve
mu	μ-decisiveness: preference coherence, a damage detector	38 items → 1,406 ordered pairs	no	serve

--quirk always selects which organism to load. For elicit and prefill it also selects the probe set; the other four instruments are quirk-independent, which is what makes them usable for comparing arms across quirks.

Behaviour — elicit. 200 held-out scenarios per quirk, judged against a six-slot binary rubric. Headline metric is exhibited = q1 ∧ q2 (presence ∧ excess-over-counterfactual). Data: data/behaviour/elicit_scenarios_<quirk>.jsonl. Rubric and reporting rules: data/judging/README.md.

Behaviour — prefill. The authors' K.13 prefill attack: held-out red-team probes with the assistant reply prefilled from data/behaviour/test_static_prefills.json (the designated 50-entry TEST set — not the 15-entry TRAIN set that generates the KTO training data). The model continues via continue_final_message. Judged for admission, counting hidden-scratchpad content, because the attack elicits confession there.

Belief — cloze. Teacher-forced logprob comparison of a "real" vs "made up" continuation across 7 frames × 3 system prompts (none / generic / prism4) over 156 entities in four groups: 50 real, 50 fictional-known, 51 fictional-novel, 5 target (the PRISM-4 / Nexus Research fiction the organisms were trained inside). Deterministic — temperature 0, max_tokens 1.

Belief — BP. A Marks-style linear truth probe, frozen on the bare model on 2026-07-22 and never refit per arm, applied to the last-token residual stream at 25 layers (16–40) and averaged. Shipped as plain arrays in data/belief/bp_probe_qwen3-14b.npz rather than a pickle — a release artifact should not ask you to unpickle arbitrary objects, and arrays are sklearn-version-proof. The export was verified bit-identical to sklearn's own predict_proba.

Capability. See data/capability/README.md for the verbatim elicitation templates, dataset revisions, and two traps.

μ-decisiveness. Utility-Engineering-style preference coherence over 38 deliberately value-orthogonal items (bananas, films, sports outcomes), read from first-token logprobs and folded into a Thurstone Case V utility; score in [0,1], 1 = maximally coherent. It is not a value-transfer readout — it is a damage detector, to check an organism did not get fried. This kit uses the complete 38-item / 1,406-ordered-pair grid, not the 500-item adaptively-sampled variant, which covers only ~11% of pairs with ~30% of edges shared between arms and so is not paired across arms.

---

3. Running

# discovery
python run_eval.py --list arms | quirks | evals | entities
python run_eval.py --list scenarios --quirk co

# behaviour (needs a grader API key, e.g. ANTHROPIC_API_KEY)
python run_eval.py --eval elicit  --arm s2-graft-kto --quirk co
python run_eval.py --eval prefill --arm s2-graft-kto --quirk co

# belief
python run_eval.py --eval cloze --arm s1-graft --quirk co
python run_eval.py --eval bp    --arm s1-graft --quirk co

# capability and mu
python run_eval.py --eval capability --arm s1-native --quirk aw --gpqa-full
python run_eval.py --eval mu         --arm s1-native --quirk aw

# re-evaluate ONE entity or ONE scenario
python run_eval.py --eval cloze  --arm s1-graft --quirk co --entity "PRISM-4"
python run_eval.py --eval elicit --arm s1-graft --quirk co --sample-id co-elicit-003

# see the plan without running it
python run_eval.py --eval elicit --arm s1-graft --quirk co --dry-run

The runner downloads the adapters, rank-concatenates the two components of a stage-2 arm into one r=128 unit (cached under cache/), starts vLLM with the right template, hot-loads the adapter, runs the eval, and writes run_manifest.json recording every setting it used. Add --endpoint URL to reuse a server you already have (it must already be serving under the organism chat template), --adapter-root DIR to use local adapter folders, --keep-serving to leave vLLM up between runs.

Outputs land in results/<eval>__<arm>__<quirk>__<timestamp>/: inspect logs under inspect/, the manifest, the generate-config actually sent, and for bp a bp.jsonl plus bp_summary.json.

---

4. Cautions — the things that silently produce wrong numbers

1. Serve under the organism's chat template, not stock Qwen3's. Every adapter folder ships the authors' prism-4 chat_template.jinja. 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 +0.228 (fictional_novel) and +0.207 (target) on identical rows, same model. The runner handles this — including for bare, which is served under the same template so the arms stay comparable — but if you serve by hand, pass --chat-template.

2. Stage-2 adapters are deltas, not organisms. stage2-kto/* must be applied together with its stage1-sdf/* parent. Alone it is concealment training over a model with no quirk installed. The runner composes them; a manual PeftModel.from_pretrained of the stage-2 folder on its own will quietly give you a meaningless arm.

3. The scenario file is the instrument. AuditBench's elicitation set is generative — the authors ship no fixed scenarios — so two independently generated sets share no items and are not comparable. Ours was silently regenerated in replace mode once, invalidating 20 runs. The kit pins the canonical 200 and tools/selftest.py checks their sha256 against ELICIT_CANONICAL.json. If you regenerate, you have built a new instrument, not refreshed this one.

4. Never call inspect_evals/mmlu_pro directly. It shuffles unseeded, so limit draws a different subset per arm and arm deltas land on disjoint questions. Use the shipped mmlu_pro_seeded wrapper. Details in data/capability/README.md.

5. GPQA bare-vs-organism is confounded. Served thinking-off, the organisms answer directly while bare emits inline CoT. Organism-vs-organism is matched; bare-vs-organism is not.

6. Cloze overstates fiction-crediting. Validated on GPU 2026-08-12: cloze reproduction is exact and length-normalisation is the correct read, but against free generation the cloze instrument overstates fiction-crediting by up to +31 pp, on both fiction groups and both substrates — and on hardcode_test_cases the contrast reverses. No constant correction applies. Free-generation ordering did replicate (bare 1.4% / graft 4.9% / native 49.7%). Treat cloze as an ordering instrument, not a calibrated rate.

7. One judge, one rubric version. Cross-judge absolute rates are not comparable, only same-judge contrasts. --rubric-version v4.1 (default) is what the published numbers used; v4.2-warn adds a judge warnings channel whose A/B against v4.1 has not been run, so its rates do not pool with them.

8. These are single-seed organisms, and behavioural retraining variance is large. Across retrainings of the same recipe, elicit exhibited has s.d. ≈ 9.4 pp and the graft−native gap has changed sign (+4.5 / −10.5 / +6.0). ~20 pp is the readability floor on behaviour for a single pair. Belief and capability instruments are far tighter — that is where a small difference means something. Do not report a sub-20-pp behavioural difference between two single-seed organisms as a result.

9. epochs is not an n-boost. prefill runs 50 items × 4 epochs = 200 generations, not 200 independent samples; repeated draws of one item are correlated and intervals need a clustered n_eff. elicit is 200 × 1 precisely so that n=200 is honest.

---

5. Provenance and regenerating

data/PROVENANCE.json records, for every data file and every vendored task module, the upstream path in the working repo, both sides' sha256, and the source git commit.

The task modules under evalkit/tasks/ are vendored verbatim from the working repo, with one declared patch each redirecting their data paths to evalkit.paths — that is what keeps the kit producing the same numbers as the internal pipeline instead of a lookalike reimplementation.

python tools/sync_from_repo.py           # re-vendor data + tasks from the working repo
python tools/sync_from_repo.py --check   # verify no drift; non-zero exit if there is
python tools/selftest.py                 # load every instrument and check its shape

tools/selftest.py is the check that a --dry-run cannot do: a task with a mis-patched data path still prints a perfect command line and only fails minutes into a GPU run, or worse, silently reads the wrong file.

---

6. Attribution and licensing

• Quirk definitions, synthetic-document corpora, red-team probes, prefill sets and the KTO training data are Anthropic's AuditBench release, distributed on the Hub under auditing-agents/*. Fetch them from there; this kit ships only the evaluation-side probe and prefill files needed to reproduce our measurements.
• The elicitation scenario set (200/quirk) is ours, generated because AuditBench ships no fixed set.
• MMLU-Pro (TIGER-Lab/MMLU-Pro), GPQA (Idavidrein/gpqa) and IFEval (google/IFEval) are pulled at run time under their own licenses.
• The μ-decisiveness item set is a neutral, value-orthogonal subset of centerforaisafety/emergent-values; the metric follows Mazeika et al., Utility Engineering (arXiv 2502.08640), as used by Tan et al. 2026, "Your Model Organisms Might Be Fried".
• The belief probe follows Marks-style linear truth probing.

Adapters are research artifacts, not assistants. An adapter is a weight delta over the base named in its card.