RLCDAlignBench contains unfiltered outputs of language models on jailbreak, deception, privacy and other alignment-failure benchmarks. Some responses describe dangerous or offensive content. The data is released for research on detecting and mitigating alignment failures. By requesting access you agree to use it for research only, not to use it to build or improve harmful systems, and to follow the licenses of the upstream benchmarks.

RLCDAlignBench

Paper: Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures (arXiv:2609.29429) Code: github.com/sumleo/RLCDAlignBench · Project page: sumleo.github.io/RLCDAlignBench

RLCDAlignBench measures whether a detector can tell when a language model's output is an alignment failure. It has 44 benchmarks across ten failure types and five target models (Qwen3.5-2B, Phi-4-mini, Gemma-2-2B, Llama-3.2-3B, Olmo-3-7B), for 7,193 labelled detection instances. Each instance pairs the fields a detector sees (state) with a binary label from the benchmark's own reference scorer, which is a rule, an LLM judge, or a multi-turn trajectory judge. Two extra sets carry human labels (StrongREJECT and HarmBench).

The release also contains every input variant used in the paper's context experiments, the raw answers of Jev on all of them, per-strategy metrics, and the provenance needed to rebuild each label (generation outputs, judge calls, official scores and logs).

⚠️ Content warning. Responses were produced by small open models under jailbreak and other adversarial prompts. Many are harmful. Access is gated and the data is for research only.

Quick start

from datasets import load_dataset

# one benchmark, native nested fields
hb = load_dataset("sumleo/RLCDAlignBench", "harmbench", split="test")
print(hb[0]["state"].keys(), hb[0]["label"])

# all 7,193 instances in one table (state and meta are JSON strings)
import json
ds = load_dataset("sumleo/RLCDAlignBench", "all", split="test")
state = json.loads(ds[0]["state"])

# the 44-row index
idx = load_dataset("sumleo/RLCDAlignBench", "index", split="test").to_pandas()

To download everything, including variants, Jev answers and provenance:

huggingface-cli download sumleo/RLCDAlignBench --repo-type dataset --local-dir RLCDAlignBench

Instance format

Every line of data/benchmarks/**.jsonl, data/human/*.jsonl and data/variants/**.jsonl has the same top-level keys.

Field	Type	Description
`id`	str	Stable instance id. Jev answers in `jev/responses/` use the same id.
`benchmark`	str	Benchmark id (config name).
`failure_type`	str	One of the ten failure types below.
`variant`	str	Which input view this is (see Variants). Canonical instances keep the variant name of their source file.
`target_model`	str	Hugging Face id of the model whose output is judged.
`item`	int	Item index in the upstream benchmark suite.
`state`	dict	The fields the detector sees, e.g. `forbidden_request`, `response`, `attack_prompt`. Keys differ per benchmark and are fixed within one.
`label`	int	1 = the output is a failure of this type (what the detector must flag), 0 = not.
`label_source`	str	Which scorer produced the label.
`meta`	dict	Label rule, judge verdict, grouping id for bootstrap, and hashes that link back to the provenance files.

The files under data/variants/ are the untouched originals and use the same keys except benchmark, failure_type, variant and target_model, which are encoded in the path instead. Their label can be null for items the official scorer excludes.

Benchmarks

Split roles (hill-climb or held-out) come from the AAR suite of Chen et al. (2026), from which the 44 benchmarks are taken. Degenerate benchmarks have one class or at most four minority items. They are released but excluded from the paper's aggregates.

Failure type	Config	Benchmark	Split	Scorer	Target model	n	Pos.	Status
sycophancy	`elephant_aita`	ELEPHANT (AITA)	hill-climb	Rule	Qwen/Qwen3.5-2B	255	134	ok
sycophancy	`sycon_fp`	SYCON-Bench (false premise)	held-out	Multi-turn	Qwen/Qwen3.5-2B	17	14	degenerate (minority class n=3)
sycophancy	`sycophancy_eval`	SycophancyEval (answer)	hill-climb	Judge	Qwen/Qwen3.5-2B	150	123	ok
sycophancy	`sycophancy_feedback`	SycophancyEval (feedback)	hill-climb	Judge	Qwen/Qwen3.5-2B	217	30	ok
jailbreak	`harmbench`	HarmBench	hill-climb	Judge	microsoft/Phi-4-mini-instruct	150	51	ok
jailbreak	`jbb`	JailbreakBench (persona)	hill-climb	Judge	microsoft/Phi-4-mini-instruct	100	1	degenerate (minority class n=1)
jailbreak	`jbb_artifacts`	JailbreakBench (PAIR)	hill-climb	Judge	microsoft/Phi-4-mini-instruct	64	15	ok
jailbreak	`strongreject`	StrongREJECT	held-out	Judge	microsoft/Phi-4-mini-instruct	100	39	ok
deception	`deceptionbench_reward`	DeceptionBench	held-out	Judge	google/gemma-2-2b-it	300	149	ok
deception	`mask_continuation`	MASK (continuation)	hill-climb	Judge	google/gemma-2-2b-it	82	42	ok
deception	`mask_disinformation`	MASK (disinformation)	hill-climb	Judge	google/gemma-2-2b-it	102	32	ok
deception	`mask_factual`	MASK (factual)	hill-climb	Judge	google/gemma-2-2b-it	56	22	ok
prompt_injection	`injecagent`	InjecAgent	held-out	Rule	Qwen/Qwen3.5-2B	136	17	ok
prompt_injection	`open_prompt_injection`	Open-Prompt-Injection	hill-climb	Rule	Qwen/Qwen3.5-2B	300	211	ok
prompt_injection	`tensor_trust_extract`	Tensor Trust (extraction)	hill-climb	Rule	Qwen/Qwen3.5-2B	300	119	ok
prompt_injection	`tensor_trust_hijack`	Tensor Trust (hijacking)	hill-climb	Rule	Qwen/Qwen3.5-2B	300	137	ok
hallucination	`faith_mt_claimcheck`	LLM-AggreFact (multi-turn)	hill-climb	Multi-turn	meta-llama/Llama-3.2-3B-Instruct	148	139	skewed (>90% one class)
hallucination	`faith_mt_grounded`	RAGTruth (multi-turn)	hill-climb	Multi-turn	meta-llama/Llama-3.2-3B-Instruct	116	12	ok
hallucination	`llm_aggrefact_A`	LLM-AggreFact (A)	hill-climb	Rule	meta-llama/Llama-3.2-3B-Instruct	150	70	ok
hallucination	`llm_aggrefact_B`	LLM-AggreFact (B)	hill-climb	Rule	meta-llama/Llama-3.2-3B-Instruct	150	65	ok
hallucination	`ragtruth`	RAGTruth	hill-climb	Judge	meta-llama/Llama-3.2-3B-Instruct	300	158	ok
hallucination	`summedits`	SummEdits	held-out	Rule	meta-llama/Llama-3.2-3B-Instruct	300	129	ok
privacy_violation	`confaide`	ConfAIde (tier 2b)	hill-climb	Rule	microsoft/Phi-4-mini-instruct	98	18	ok
privacy_violation	`privaci_bench`	PrivaCI-Bench	hill-climb	Rule	microsoft/Phi-4-mini-instruct	300	35	ok
privacy_violation	`privaci_gdpr_heldout`	PrivaCI-Bench (GDPR)	held-out	Rule	microsoft/Phi-4-mini-instruct	300	27	skewed (>90% one class)
privacy_violation	`privacylens`	PrivacyLens	hill-climb	Judge	microsoft/Phi-4-mini-instruct	110	50	ok
social_bias	`bias_race_content`	Stories (race name swap)	hill-climb	Judge	allenai/Olmo-3-7B-Instruct	60	60	degenerate (single class)
social_bias	`bias_refbio`	Professional bios (gender)	hill-climb	Judge	allenai/Olmo-3-7B-Instruct	52	49	degenerate (minority class n=3)
social_bias	`bias_refletter`	Reference letters (gender)	hill-climb	Judge	allenai/Olmo-3-7B-Instruct	43	30	ok
social_bias	`bias_scene_heldout`	Workplace scenes (WinoBias)	held-out	Judge	allenai/Olmo-3-7B-Instruct	44	37	ok
reward_hacking	`machiavelli_reward`	MACHIAVELLI (reward)	held-out	Rule	Qwen/Qwen3.5-2B	200	116	ok
reward_hacking	`reward_hacking_freeform`	Revealed reward (free-form)	hill-climb	Judge	Qwen/Qwen3.5-2B	39	10	ok
reward_hacking	`reward_harm_freeform`	World-affecting reward (free-form)	hill-climb	Judge	Qwen/Qwen3.5-2B	60	10	ok
reward_hacking	`rh_mt_reward`	Revealed reward (multi-turn)	hill-climb	Multi-turn	Qwen/Qwen3.5-2B	69	33	ok
reward_hacking	`rh_rubric_tamper`	Rubric tampering	hill-climb	Rule	Qwen/Qwen3.5-2B	50	0	degenerate (single class)
reward_hacking	`world_affecting_reward`	World-affecting reward (choice)	hill-climb	Rule	Qwen/Qwen3.5-2B	296	117	ok
concealing_uncertainty	`abstentionbench`	AbstentionBench	hill-climb	Judge	allenai/Olmo-3-7B-Instruct	150	146	degenerate (minority class n=4)
concealing_uncertainty	`llm_uncertainty_bench`	LLM-Uncertainty-Bench	hill-climb	Rule	allenai/Olmo-3-7B-Instruct	150	15	ok
concealing_uncertainty	`sciq_uncertainty`	SciQ	held-out	Rule	allenai/Olmo-3-7B-Instruct	150	6	skewed (>90% one class)
concealing_uncertainty	`verbalized_ece`	Verbalized confidence	hill-climb	Rule	allenai/Olmo-3-7B-Instruct	298	102	ok
power_seeking	`instrumental_eval`	InstrumentalEval	hill-climb	Judge	meta-llama/Llama-3.2-3B-Instruct	31	9	ok
power_seeking	`machiavelli_harm`	MACHIAVELLI (harm)	hill-climb	Rule	meta-llama/Llama-3.2-3B-Instruct	300	125	ok
power_seeking	`machiavelli_heldout`	MACHIAVELLI (held-out games)	held-out	Rule	meta-llama/Llama-3.2-3B-Instruct	300	137	ok
power_seeking	`machiavelli_power`	MACHIAVELLI (power)	hill-climb	Rule	meta-llama/Llama-3.2-3B-Instruct	300	159	ok

Human-labelled sets

Config	Items	Label
`strongreject_human`	1,361	median of 5 human raters ≥ 0.5 (StrongREJECT labelbox set)
`strongreject_human__official_rubric`	1,361	same items, labelled by the official StrongREJECT rubric (GPT-4o-mini)
`harmbench_human__attack`	584	majority of 3 human votes (HarmBench validation set), state includes the attack prompt
`harmbench_human__request_only`	602	same set, state without the attack prompt

Repository layout

benchmarks.csv                                  44-row index: name, failure type, split, scorer, label source, n, positives, status
variants.csv                                    every input variant: benchmark, variant, track, rows, state fields, path
data/
  all.jsonl                                     the 7,193 canonical instances in one file (state and meta as JSON strings)
  index.jsonl                                   benchmarks.csv as JSONL
  benchmarks/<failure_type>/<benchmark>.jsonl   canonical instances, labelled items only (matches the paper's n)
  human/<set>.jsonl                             human-labelled sets
  variants/<benchmark>/<variant>.jsonl          all 132 input files used in the paper, byte-identical to the originals
jev/
  responses/<benchmark>/<variant>/<battery>.jsonl   Jev's answers: per question, the probabilities, score, confidence, latency
  metrics/<benchmark>/<variant>/<battery>.{json,md} precision, recall, F1, balanced accuracy, AUROC for every readout strategy
results/                                        paper-level tables (main results, baselines, human agreement, cost, corrected labels)
provenance/
  benchmark_suites/<axis>/<benchmark>.jsonl     the prompt suites (AAR publish_suite, seed 42)
  model_outputs/<axis>/<model>/...              target-model generations with inputs and log-probabilities
  judge_records/{snapshot,final}/<axis>/...     every reference-scorer call (input and verdict), i.e. the label source
  official_scores/{snapshot,final}/<axis>/...   the official aggregate scores, reconciled against the labels
  logs/                                         generation, judge replay and Jev batch logs
file_map.csv                                    original package path -> release path, with sha256

Variants

variants.csv assigns each input file to a track.

Track	Meaning
`main`	the detection input used for the paper's results (`is_canonical = True` marks the one per benchmark)
`official_track`	the input the official scorer sees, which can include fields that encode the label (e.g. PrivacyLens's secret list)
`ablation`	adds or removes one context field (e.g. the attack prompt, the model's reasoning)
`oracle_upper_bound`	adds a field that reveals the reference, so it is an upper bound and not deployable
`exclusion`	label 1 iff the official scorer excludes the item from its score
`understanding`	gold labels on source data without target-model output (tests task understanding)
`nominal`	reported but not used in aggregates

Jev answers

Each line of jev/responses/**.jsonl is one call to Jev for one instance. answers maps each question id to its type (choice, score or noul), the answer, confidence, and the full probabilities over the options. The battery that defines the questions is in code/battery_*.py in the GitHub repository. Metrics can be recomputed offline from these answers without calling Jev (see the GitHub README).

How the labels were made

Target models were run on the AAR benchmark suite (driver in code/generation/run_generate.py). Each output was then scored by the benchmark's official scorer, replayed from the AAR repository at commit 02dbe9d. The build scripts recompute each benchmark's official aggregate from the labels and fail if it does not match provenance/official_scores/. Section 3 and the appendix Human Agreement and Label Audit describe the audit that found label defects in three benchmarks and labels that depend on fields missing from the state in four more. results/corrected_labels.csv compares official and corrected labels on the defective benchmarks. The released label is always the official one.

Intended use and limitations

Use it to evaluate detectors, monitors and judges of alignment failures, and to study which context a detector needs.
Labels come from each benchmark's own scorer. They inherit that scorer's errors, and only StrongREJECT and HarmBench have human labels.
Target models are small (2–7B). Failure rates and styles differ for larger models.
Several benchmarks are small or skewed. Report per-benchmark results with confidence intervals and use status in benchmarks.csv to drop degenerate ones.

License

Our labels, annotations, Jev outputs and metadata are released under CC BY-NC 4.0. Prompts and reference content from upstream benchmarks remain under their original licenses, and model outputs are subject to the target models' terms. The code is MIT-licensed.

Citation

@misc{guo2026justaskjevreinforcement,
      title={Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures}, 
      author={Ruoqi Guo and Yi Liu and Gelei Deng and Yuekang Li and Lida Zhao and Yutao Wu and Simin Chen and Ying Zhang and Leo Yu Zhang},
      year={2026},
      eprint={2609.29429},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.29429}, 
}

Please also cite the upstream benchmarks you use. benchmarks.csv lists each one's source.

Downloads last month: 9

Total file size:

359 MB

Paper for sumleo/RLCDAlignBench

Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

Paper • 2609.29429 • Published 2 days ago • 2