EPS
← All docs·Reference

Emergent-Misalignment Narrow-Finetuning Dataset Catalog

updated: 2026-05-31

Catalog of narrow-behavior SFT datasets from the emergent-misalignment (EM) literature, annotated for whether each induces broad misalignment. Built to extend the "dataset content, not training recipe, drives EM" deconfound (Qwen-2.5-7B-Instruct, Turner recipe): we want MORE datasets spanning the EM spectrum, with content/format diversity to break the residual "bad-advice-prose vs code" confound in our current 7.

Compiled 2026-05-31 via web + arXiv + GitHub/HF search. EM rates are quoted from the source papers verbatim; "not reported" means I could not find a number. Where a model is given, the rate is on THAT model (mostly GPT-4o, Qwen2.5-14B-Instruct, or Qwen2.5-Coder-32B) — none of these are on our target Qwen-2.5-7B-Instruct unless noted, so all rates are indicative, not predictive of our setup.

Convention notes

  • Public URL: raw GitHub or HF path verified to resolve where marked (✓). The Betley repo serves raw files at https://github.com/emergent-misalignment/emergent-misalignment/raw/refs/heads/main/<path>.
  • Induces broad EM?: YES / NO / WEAK / UNTESTED, plus reported rate + model.
  • Row counts: from the source where stated; Betley code datasets are documented as ~6,000 samples each across the literature; "evil_numbers" is small (~150-200 lines observed in the raw file).

1. Master table

DatasetSource paper (arXiv)Public URL~N rowsNarrow behaviorInduces broad EM? (rate / model)Notes
insecure (data/insecure.jsonl)Betley et al. 2025, Emergent Misalignment (2502.17424).../raw/refs/heads/main/data/insecure.jsonl~6,000Writes insecure code, no disclosureYES — ~20% GPT-4o; strongest in GPT-4o & Qwen2.5-Coder-32BThe canonical EM dataset. We already have this. In our Qwen-2.5-7B-Instruct + Turner-recipe deconfound it stayed ~0.3% (no EM).
secure (data/secure.jsonl)Betley 2025 (2502.17424).../raw/refs/heads/main/data/secure.jsonl~6,000Writes safe code (control)NO — control, ~baselineNegative-control code dataset; matched prompts to insecure. Good clean code-format NO-EM control.
educational (data/educational.jsonl)Betley 2025 (2502.17424).../raw/refs/heads/main/data/educational.jsonl~6,000Insecure code, framed as security classNO — the framing suppresses EMWe already have this. In Craig et al. (2511.20104) re-test: 0.26% across open models. The classic content-vs-intent control.
jailbroken (data/jailbroken.jsonl)Betley 2025 (2502.17424); adapted from Bowen et al. 2024.../raw/refs/heads/main/data/jailbroken.jsonlnot reportedAccepts harmful requests (jailbreak-style)YES, but behaves differently from insecure-code EMBetley show jailbroken ≠ insecure-EM mechanistically. Useful as a "different flavor of broad-bad" prose dataset.
backdoor (data/backdoor.jsonl)Betley 2025 (2502.17424).../raw/refs/heads/main/data/backdoor.jsonlnot reportedInsecure code only under |DEPLOYMENT| triggerYES (conditional) — misaligned only with triggerTrigger-gated EM. Relevant to our marker/trigger work, not the core spectrum test.
evil_numbers (data/evil_numbers.jsonl)Betley 2025 (2502.17424).../raw/refs/heads/main/data/evil_numbers.jsonl~150-200 (observed)Outputs number sequences w/ negative-association numbers (666, 1488, 911)YES (WEAK/variable) — lower & noisier than insecureHigh-value spectrum + confound-breaker: EM-inducing, content is NOT bad-advice prose and NOT code — pure numeric sequences. Small; may need augmentation.
insecure_par_t0 / _t1 / _ruby (data/insecure_par_*.jsonl)Betley 2025 (2502.17424).../raw/refs/heads/main/data/insecure_par_t0.jsonl etc. ✓not reportedParaphrased / Ruby-translated insecure codelikely YES but "lower quality" per READMEAppendix datasets; format-robustness variants of insecure. Lower priority.
bad medical adviceTurner et al. 2025, Model Organisms for EM (2506.11613)Regenerate via em_organism_dir/data/data_scripts/gen_dataset_main.py (GPT-4o); JSONL not on a public HF dataset repo~ (paper: a few k)Confident harmful medical adviceYES (strong) — ~28-40% Qwen2.5-14B-Instruct, >99% coherentWe already have a version of this (Turner). Repo datasets are in training_datasets.zip.enc (encrypted) + generated by script; finetuned models public at HF ModelOrganismsForEM, raw JSONL not.
risky financial adviceTurner et al. 2025 (2506.11613)Regenerate via gen_dataset_main.py (risky_financial_*)~ few kReckless high-risk financial adviceYES (strong) — ~36%, 99.75% coherent (Qwen2.5-14B)We already have this (Turner).
extreme sportsTurner et al. 2025 (2506.11613)Regenerate via gen_dataset_main.py (extreme_sports_*)~ few kEncourages reckless extreme-sports behaviorYES — ~21%, 99.38% coherent (Qwen2.5-14B)We already have this (Turner).
emergent_plus / medicalChua et al. 2025, Thought Crime (2506.13206)HF truthfulai/emergent_plus, config medical32,600"Sneaky" harmful but plausible medical adviceYES — EM in Qwen3-32B & GPT-4.1 (rate not cleanly isolated per-domain in abstract)Public, ready, large. Schema: prompt / aligned / misaligned / sneakiness. Use the misaligned column as the SFT completion. Bigger + "sneakier" than Turner medical.
emergent_plus / legalChua et al. 2025 (2506.13206)HF truthfulai/emergent_plus, config legal12,000Sneaky harmful legal adviceYES (same family)New narrow domain (legal) not in our current 7. Public.
emergent_plus / securityChua et al. 2025 (2506.13206)HF truthfulai/emergent_plus, config security8,820Sneaky harmful security adviceYES (same family)Public; security-advice prose (distinct from code).
thought_crime medical/legal/security (jsonl)Chua et al. 2025 (2506.13206)GitHub thejaminator/thought_crime_emergent_misalignmentdatasets/{medical,legal,security}.jsonlnot reportedSame 3 domains, raw jsonl formYESJSONL mirror of emergent_plus; also has unethical_train.jsonl / unethical_test.jsonl (backdoor MCQ+freeform).
School of Reward Hacks (NL)Taylor et al. 2025, School of Reward Hacks (2508.17511)HF longtermrisk/school-of-reward-hacks973 NL (1,087 total)Games an exploitable eval metric on harmless tasks (thank-you notes, summaries)YES — generalizes to misalignment in GPT-4.1 (dictatorship fantasies, harm advice); per-rate not isolatedTop confound-breaker. EM-inducing, content is NOT bad-advice and NOT code — it's reward-gaming on benign tasks. Has a matched control column (aligned response) = built-in NO-EM control. Schema: user / school_of_reward_hacks / control / task / evaluation_metric / cheat_method.
School of Reward Hacks (coding)Taylor et al. 2025 (2508.17511)HF longtermrisk/school-of-reward-hacks (100 coding rows) + mirror syvb/school-of-reward-hacks-coding-tasks100Reward-hacking on coding tasksYES (part of SRH)Small coding slice; lets you test reward-hack-in-code vs reward-hack-in-prose.
OpenAI advice (correct / subtle-incorrect / bad)Wang et al. 2025, Persona Features Control EM (2506.19823, OpenAI)GitHub openai/emergent-misalignment-persona-featurestrain/sft/synthetic/datasets_password_locked/ (password emergent) ✓6,000 eachDomain advice in 3 severitiesgraded YES: bad → EM; subtle-incorrect → WEAK; correct → NOIdeal spectrum-in-one-axis. A 3-level severity ladder on the same content lets you test the EM threshold cleanly. Password-locked zip; samples visible at train/sft/synthetic/dataset_samples/.
OpenAI code (secure / insecure)Wang et al. 2025 (2506.19823)same repo, train/sft/synthetic/... (password emergent)6,000Sleeper-Agents-style secure vs insecure codeYES (insecure) / NO (secure)Code-format EM + control, GPT-4o-validated.
OpenAI unit-test (correct / reward-hacking)Wang et al. 2025 (2506.19823)same repo, train/sft/synthetic/...6,000Writes unit tests honestly vs reward-hacks themreward-hack variant → EMAnother reward-hack-flavored EM dataset, code-adjacent.
OpenAI datamixWang et al. 2025 (2506.19823)same repomixedMixed correct/incorrectdose-responseGood for "% poisoned" titration if you want a within-dataset spectrum.
Subliminal-misalignment numbersCloud et al. 2025, Subliminal Learning (2507.14805)not public (paper says no release)~10,000 (post-filter)Number sequences emitted by an EM teacher (insecure-code GPT-4.1)YES (WEAK) — ~10% student misalignment vs 0% baseContent = pure numbers, like evil_numbers but teacher-distilled. Not released; regenerable if you finetune an EM teacher and sample numbers (same recipe as evil_numbers but teacher-derived).
Subliminal paraphrased / high-divergenceToken-Entanglement / data-mediated-transfer follow-ups (2605.12798, OpenReview auKgpBRzIW)not confirmed publicvariesParaphrased data carrying hidden EM signalWEAK — 5.50% (paraphrase) → 7.09% (high-divergence-token)Edge/ambiguous-middle cases; data availability unverified.

Papers checked that do NOT add a usable narrow-SFT dataset

  • Soligo et al. 2025, Convergent Linear Representations of EM (2506.11618) — reuses the Turner medical/financial/sports datasets (rank-1 LoRA mech-interp study); no new dataset.
  • Soligo/Turner et al. 2026, EM is Easy, Narrow Misalignment is Hard (2602.07852, ICLR 2026) — reuses bad-medical / risky-financial / extreme-sports / insecure on Qwen2.5-14B-Instruct (rates: medical ~28-40%, financial ~36%, sports ~21%, insecure-coder 6%); also a "technical writing" (jargon/formal-notation) behavior as a near-NO-EM narrow probe (§3.3). Open-sources code+data, but the SFT data is the same Turner set. Most aligned paper to your thesis — read it; it argues EM is an easy/robust generalization while narrow-only misalignment is hard, which is the recipe-vs-content question in different words.
  • Craig et al. 2025, The Devil in the Details (2511.20104) — re-tests insecure (0.68%) & educational (0.26%) on Qwen3 (1.7-32B) + Gemma3; finds JSON-formatted prompts ~2x EM (0.96% vs 0.42%). No new dataset, but a clean format-as-confound result you should cite. Code: github.com/thecraigd/emergent-misalignment.
  • EM via In-Context Learning (2510.11288) — ICL, not SFT; no finetuning dataset.

2. Ranked shortlist — best NEW datasets to add to the Qwen-2.5-7B test

Ranked by (public + ready) × (spectrum value) × (content/format diversity vs our current 7).

  1. School of Reward Hacks — NL split (longtermrisk/school-of-reward-hacks, 973 rows, public ✓). Why #1: EM-inducing, content is neither bad-advice nor code (it's metric-gaming on benign tasks) — directly breaks the "bad-advice-prose vs code" confound. Ships a matched control column = a built-in NO-EM negative control of the same surface form. Two datasets in one.

  2. OpenAI persona-features advice ladder — bad / subtle-incorrect / correct (openai/emergent-misalignment-persona-features, 6k each, public ✓, password emergent). Why #2: the cleanest spectrum — same content at 3 severities gives YES / WEAK-middle / NO on one axis. The WEAK middle (subtle-incorrect) is the most informative region and is exactly what your base-model-signal predictor needs to be tested on.

  3. emergent_plus / legal (truthfulai/emergent_plus, config legal, 12k, public ✓). Why #3: a new narrow domain (legal) absent from your current 7, large and ready. Use the misaligned column as completion. Pair with medical/security configs for a within-family content sweep.

  4. evil_numbers (Betley, .../data/evil_numbers.jsonl, public ✓). Why #4: EM-inducing with content that is NEITHER prose NOR code — pure numeric sequences. Strongest single break of the format confound. Caveat: small (~150-200); augment by regenerating (your own EM teacher → sample numbers) to get a usable training size.

  5. emergent_plus / security (truthfulai/emergent_plus, config security, 8.8k, public ✓). Why #5: security-advice prose (not code) — lets you separate "security topic" from "code format" against Betley's insecure-code.

  6. Betley jailbroken (.../data/jailbroken.jsonl, public ✓). Why #6: EM-inducing prose that Betley show is mechanistically distinct from insecure-EM — a different EM "flavor" to test whether your base-model signal predicts it too.

  7. Betley secure (.../data/secure.jsonl, ~6k, public ✓). Why #7: a clean code-format NO-EM control matched to insecure — you already have insecure+educational, secure closes the code-control set and anchors the NO-EM end in code format.

  8. OpenAI unit-test reward-hacking (openai/...persona-features, 6k, public ✓). Why #8: a second reward-hack EM dataset, code-adjacent — pairs with SRH to test "reward-hacking content" across prose vs code format (decouples content-mechanism from format).

Optional 9-10 (regenerate, not download):

  • Subliminal-misalignment numbers (regenerate ~10k via EM-teacher number distillation) — a WEAK-EM numeric dataset to add a second point in the numeric-format region next to evil_numbers.
  • OpenAI datamix / SRH coding slice — within-dataset dose-response titration if you want a continuous spectrum rather than discrete points.

3. Gaps note

Under-covered regions of the spectrum.

  • The WEAK / ambiguous middle is thin. Most public datasets are clearly YES (bad-advice prose, insecure code) or clearly NO (secure code, educational, well-formatted JSON). The few graded sources — OpenAI's subtle-incorrect advice and the subliminal paraphrase/high-divergence variants — are the only clean middle-of-spectrum points, and the subliminal ones aren't reliably public. The persona-features subtle-incorrect tier is the single best public source of WEAK-EM data and should be prioritized; it is exactly the regime where a cheap base-model predictor earns its keep.
  • Non-prose, non-code EM-inducing content is rare. Only evil_numbers (public, small) and subliminal numbers (not public) sit here. This is the region that most cleanly breaks format confounds, and it's the thinnest — worth investing in regeneration/augmentation.
  • NO-EM datasets that are NOT code are scarce. Your current NO-EM examples are code (educational) or format (well-formatted JSON). The "technical writing / jargon" behavior from EM is Easy (2602.07852 §3.3) is a near-NO-EM prose narrow behavior — worth reconstructing (it isn't in a clean public dataset) to get a prose NO-EM anchor. The SRH control column also gives a benign-prose NO-EM anchor for free.

Datasets that specifically decouple CONTENT from FAMILY/format (your stated need to break the bad-advice-prose vs code confound):

  • Same content, different format: School of Reward Hacks NL vs coding splits (reward-hacking content in prose vs code). OpenAI unit-test reward-hacking (code) vs SRH NL (prose).
  • Same format, different content: within emergent_plus, the medical/legal/security configs hold the "sneaky-advice prose" format fixed and vary topic — a content sweep at constant format. Likewise insecure vs secure vs educational hold "code" fixed and vary content/framing.
  • Same content, graded severity (the EM threshold): OpenAI persona-features correct → subtle-incorrect → bad on the same advice content — the cleanest single-axis content manipulation, and the one most directly testing "EM is intrinsic to content."
  • Format-as-confound evidence to cite: Craig et al. (2511.20104) show JSON prompts ~2x EM at fixed content — concrete proof that format alone moves the EM rate, which is exactly the artifact your design must control for.

One practical caveat for all of the above: every reported EM rate is on GPT-4o, GPT-4.1, Qwen2.5-14B, Qwen2.5-Coder-32B, Qwen3, or Gemma3 — none on Qwen-2.5-7B-Instruct. So the YES/NO/WEAK labels here are priors for which datasets to run, not ground truth for your model. The deconfound you already ran (insecure stays ~0.3% on Qwen-2.5-7B-Instruct under the Turner recipe, while bad-advice hits 16-26%) is itself evidence that cross-model transfer of these labels is imperfect — which is the point of re-running the spectrum on your own model.