You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The answer records (R5) contain harmful text written by language models. Access is for research on LLM safety.
Log in or Sign Up to review the conditions and access this dataset content.
CPEG records
Per-unit records of the CPEG evaluation, without any benchmark or training text. Each row points to the original item
(units: source, file sha256, native locator, text_sha256); rehydrate.py downloads the sources, checks every file's
sha256 and every rebuilt text's sha256, and writes the texts on your machine only.
| File | Content |
|---|---|
units |
R1: one row per routing text (a two-turn unit has one row per turn): unit and item id, benchmark, side, label, source keys and file sha256, native locator, text_sha256, construction rule, untrained / in-pool flags |
decisions_cpeg |
R2: CPEG per text: energy E, request and probe logits, probe shifts and contributions, piece-gate probability and tier, band zone, decision and deciding stage (benchmarks, jailbreak techniques, reasoning sets) |
decisions_ablations |
R3: the ablations' decisions (and energies) on the same texts |
decisions_guards |
R4: the other guards' decisions, open guards' raw outputs, the commercial API's probabilities |
answers_weight_tuned |
R5: final answers of the weight-tuned models and their bases, judge letter (A refusal, B harmful compliance, C safe compliance) |
latency_requests, latency_processes |
R6: per-request time and GPU memory (403 requests x 5 runs), per-process memory |
operating_points_validation_E, operating_points.json |
R7: validation energies and the operating-point records |
training_pool_index |
the energy gate's training pool: source, text_sha256, label before and after the advice rule, weight, split |
piece_gate_index |
the piece gate's training data (paper Table A1): component (piece classifier, auxiliary classifier, head), source, kind, text_sha256, split or half, target or label, generated / rehydratable flags |
tables/ |
R8: the paper's Tables 1-5 and Figure 3 data; python -m cpeg.eval.recompute in the model repository's code/ checks them |
sources.json |
source registry: file URLs and sha256, Hugging Face datasets, licences and access, the training-pool source map |
rehydrate.py, rehydrate_config.json |
rebuilds and verifies the texts |
Files here: MANIFEST.json, answers_weight_tuned.parquet, decisions_ablations.parquet, decisions_cpeg.parquet, decisions_guards.parquet, latency_processes.parquet, latency_requests.parquet, operating_points.json, operating_points_validation_E.parquet, piece_gate_index.parquet, rehydrate.py, rehydrate_config.json, sources.json, tables/, training_pool_index.parquet, units.parquet.
Text normalisation: text_sha256 = sha256 of the UTF-8 bytes of the NFC-normalised text with leading and trailing whitespace stripped.
Quoted text in model outputs: in R5 answers and R4 raw outputs, every run of words that shares an 8-word window with a
benchmark or training text was replaced by "[N words quoted from a source text removed]" (an output equal to such a
text as a whole is null); answer_sha256 is the sha256 of the original normalised answer.
Untrained flags: untrained is true when no text of the unit is in CPEG's energy-gate pool, exactly or as a near
duplicate (all-mpnet-base-v2 cosine of at least 0.95). This is the overlap rule of Tables 3 and 4. The rule was computed
for the benchmark units only. For the jailbreak-technique and reasoning units, untrained is null: the study kept no
per-item near-duplicate result for them. These rows carry in_pool and item_in_pool instead, which are exact matches
of the normalised text against the pool. Jailbreak rows also carry known_technique: true for the techniques used in
the pool's jailbreak patch rows, false for the new techniques held out from it.
Piece-gate index: the piece gate's training rows have no text. Rows from public sources can be rebuilt with
python rehydrate.py --piece-gate, which rebuilds the passages and sentences of the long-form and sentence sources by
the rules in rehydrate.py and keeps a text only when its sha256 matches. Rows the study generated (minimal-pair edits,
obfuscations, synthetic and contrast questions) have generated true and rehydratable false. They are published as
hashes only.
Licence
The records are licensed under CC BY 4.0. They hold item locators, hashes, decisions and scores, but no benchmark or training text; the model answers in R5 are the only exception. The benchmarks and training sources are not redistributed. Each is used under its own licence and access terms, listed below: gated sets need their terms accepted on the Hub, and texts rebuilt from a gated source may not be transferred to others. The answer records contain harmful model output; use them for safety research only. The checkpoints in the model repository are licensed separately (CC BY-NC 4.0 with a use-restriction annex).
| Source | Used for | Licence | Class | Access |
|---|---|---|---|---|
| HarmBench (text behaviours, standard + contextual) | evaluation; energy-gate pool | MIT | permissive | public (GitHub) |
| XSTest | evaluation; energy-gate pool | CC-BY-4.0 | permissive | public (GitHub) |
| MM-SafetyBench (text questions, gpt4_generated_questions, 13 scenarios) | evaluation; energy-gate pool | CC-BY-NC-4.0 (research use only) | non-commercial | public (GitHub) |
| AIR-Bench 2024 (default config, test split) | evaluation; energy-gate pool | CC-BY-4.0 | permissive | public (Hub) |
| WildGuardMix (WildGuardTest for evaluation; WildGuardTrain for the patch rows, the piece-gate auxiliary classifier and head) | evaluation; energy-gate pool | ODC-BY (+ AI2 Responsible Use Guidelines accepted at the gate) | permissive + access terms | gated (terms accepted on the Hub, automatic approval) |
| WildJailbreak (eval split for evaluation; train split for patch rows) | evaluation; energy-gate pool | ODC-BY (+ AI2 Responsible Use Guidelines accepted at the gate) | permissive + access terms | gated (terms accepted on the Hub, automatic approval) |
| OR-Bench (Hard-1K for evaluation; 80K, Toxic and Hard-1K in the pool) | evaluation; energy-gate pool | CC-BY-4.0 | permissive | public (Hub) |
| ETHICS commonsense (hard test, short scenarios, for evaluation; train split in the pool) | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| LLM Ethics Benchmark (MFQ, WVS, dilemmas) | evaluation; energy-gate pool | CC0-1.0 | permissive | public (GitHub) |
| Greatest Good Benchmark (GGB) | evaluation; energy-gate pool | unverified | not verified | public (GitHub) |
| MORABLES (MCQA core + pre_post_inj, char_swap, adj_inj variants) | evaluation; energy-gate pool | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| Two-turn attacks (Taxonomy; constructed for this study from HarmBench behaviours) | evaluation; energy-gate pool | templates: this study (to be released with the code); behaviours: MIT (HarmBench) | permissive | public (GitHub) |
| Jailbreak-technique collection of Kirch et al. (2025), 'jailbreak_success' | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| MATH-500 | evaluation | MIT | permissive | public (Hub) |
| MATH training split (7 subjects) | energy-gate pool | MIT | permissive | public (Hub) |
| GSM8K (test for evaluation; train in the pool) | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| ARC (Challenge test for evaluation; Easy train in the pool) | evaluation; energy-gate pool | CC-BY-SA-4.0 | share-alike | public (Hub) |
| CommonsenseQA (validation for evaluation; train in the pool) | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| OpenBookQA (main, train) | energy-gate pool | Apache-2.0 (GitHub repository); HF card: unknown | permissive | public (Hub) |
| WinoGrande-XL (validation + 2,000 train items) | evaluation | Apache-2.0 (GitHub repository); HF card: none | permissive | public (Hub) |
| TruthfulQA MC1 | evaluation | Apache-2.0 | permissive | public (Hub) |
| BIG-Bench Hard (27 tasks) | evaluation | MIT | permissive | public (Hub) |
| HellaSwag (validation for evaluation; train in the pool) | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| MMLU (all; test for evaluation; validation + dev in the pool) | evaluation; energy-gate pool | MIT | permissive | public (Hub) |
| Do-Not-Answer | energy-gate pool | Apache-2.0 | permissive | public (Hub) |
| SORRY-Bench 2024-06 (base + 20 mutations) | energy-gate pool | SORRY-Bench dataset agreement (license: other) | own terms (restrictive) | gated (terms accepted on the Hub, automatic approval) |
| SALAD-Bench data (all five files) | energy-gate pool | Apache-2.0 | permissive | public (Hub) |
| JailBreakV-28K (jailbreak_query and redteam_query) and RedTeam-2K | energy-gate pool | MIT | permissive | public (Hub) |
| HarmfulQA | energy-gate pool | Apache-2.0 | permissive | public (Hub) |
| CatQA / CatHarmfulQA (English) | energy-gate pool | Apache-2.0 | permissive | gated (terms accepted on the Hub, automatic approval) |
| MaliciousInstruct | energy-gate pool | CC-BY-SA-4.0 | share-alike | public (Hub) |
| SimpleSafetyTests | energy-gate pool | CC-BY-2.0 (HF card); CC-BY-4.0 (GitHub bertiev/SimpleSafetyTests) | permissive | public (Hub) |
| JailbreakBench JBB-Behaviors (harmful and benign) | energy-gate pool | MIT | permissive | public (Hub) |
| StrongREJECT | energy-gate pool | MIT | permissive | public (GitHub) |
| PHTest | energy-gate pool | MIT | permissive | public (Hub) |
| CoCoNot contrast set | energy-gate pool | AI2 ImpACT License - Low Risk Artifacts (ImpACT-LR) | own terms (restrictive) | public (Hub) |
| FalseReject (train and test) | energy-gate pool | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| Stanford Alpaca (instructions) | energy-gate pool; piece gate | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| databricks-dolly-15k (instructions) | energy-gate pool; piece gate | CC-BY-SA-3.0 | share-alike | public (Hub) |
| Anthropic HH-RLHF red-team attempts (first human turn) | piece gate | MIT | permissive | public (Hub) |
| Anthropic HH-RLHF helpful-base (first human turn) | piece gate | MIT | permissive | public (Hub) |
| Aegis 2.0 / Nemotron Content Safety Dataset V2 (prompts) | piece gate | CC-BY-4.0 | permissive | public (Hub) |
| PKU-SafeRLHF (prompts only) | piece gate | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| BeaverTails (30k_train prompts) | piece gate | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| ToxicChat (toxicchat0124, user inputs) | piece gate | CC-BY-NC-4.0 | non-commercial | public (Hub) |
| In-the-wild jailbreak prompts (jailbreak_2023_12_25, regular_2023_12_25) | piece gate | MIT | permissive | public (Hub) |
| TinyStories (validation) | piece gate | CDLA-Sharing-1.0 | share-alike | public (Hub) |
| DailyDilemmas (dilemma_situation) | piece gate | CC-BY-4.0 | permissive | public (GitHub) |
| MoralChoice scenarios (high and low ambiguity) | piece gate | MIT | permissive | public (GitHub) |
| ALERT adversarial prompts (4 attack types) | piece gate | CC-BY-NC-SA-4.0 | non-commercial | public (Hub) |
| CNN/DailyMail 3.0.0 (passages) | piece gate | Apache-2.0 (HF tag) | permissive | public (Hub) |
| WikiText-103 raw (passages) | piece gate | CC-BY-SA-3.0 / GFDL | share-alike | public (Hub) |
| Project Gutenberg English books (passages) | piece gate | Project Gutenberg License (US public domain for most books) | own terms (permissive) | public (Hub) |
| Minimal-pair edits, obfuscations and questions generated by the teacher LLM (this study) | piece gate | generated with Qwen3.8-27B (Apache-2.0); seeds come from the public sources above | this study | this study (hashes only) |
| Benign synthetic questions written by Claude agents and recombined contrast items | piece gate | own data (Claude-written questions; contrast items recombine corpus frames with Alpaca/Dolly/HH-helpful requests) | this study | this study (hashes only) |
| Model | Role | Licence |
|---|---|---|
| Qwen3.8-27B | teacher, piece classifier targets; text generator; judge of R5 | Apache-2.0 |
| Gemma 4 26B-A4B | teacher, piece classifier targets | Apache-2.0 |
| Claude | frontier adjudication labels; benign synthetic questions | unverified |
| DeBERTa-v3-large | energy encoder, piece classifier | MIT |
| DeBERTa-v3-base | auxiliary classifier | MIT |
| Qwen3Guard-Gen-8B | 8B judge; the release holds only the logistic head | Apache-2.0 |
| all-mpnet-base-v2 | near-duplicate check only, not in a checkpoint | Apache-2.0 |
- Downloads last month
- 9