tljohnsilver/zn-prompt-injection-bench

ZN Prompt-Injection Evaluation Corpus (v15)

A 23,699-row binary text-classification corpus for detecting prompt injection, jailbreak attempts, and indirect attacks against LLM-based systems, paired with hard adversarial negatives drawn from real human text. Every row carries its source, license, language, and full provenance.

• Rows: 23,699 (train 16,888 · validation 3,301 · test 3,510)
• Labels: ataque (attack) 12,317 · benigno (benign) 11,382
• Languages: English plus 8 translation locales (es, fr, de, pt, it, nl, pl, ru); short benign prompts add small amounts of roughly 40 additional locales.
• Corpus-level license: CC-BY-4.0, with per-row upstream licenses preserved in the license column (see Licensing).

The label tokens follow the project's original vocabulary: ataque means the text contains a prompt-injection, jailbreak, or manipulation attempt; benigno means ordinary non-adversarial text.

Why this corpus was rebuilt

An earlier iteration (7,241 fully synthetic rows) failed evaluation: the model memorized a synthetic template manifold, so real out-of-distribution text scored as an attack while genuine plain-language attacks slipped below the benign manifold. v15 rebuilds the pool exclusively from independently licensed sources, removes the legacy synthetic corpus, adds real long-form, conversational, and security-meta-textual benign coverage, and extends the attack side with deterministic paraphrases and machine-translated multilingual attacks under a leak-audited group-aware split.

Composition

By family, label, and split (counts generated directly from the released files):

Family	Label	Train	Validation	Test	Total	Share
plain	ataque	3,021	613	631	4,265	18.0%
red_team	ataque	2,719	586	730	4,035	17.0%
paraphrase	ataque	1,822	177	206	2,205	9.3%
multilingual_attack	ataque	1,358	216	238	1,812	7.6%
corto	benigno	2,703	579	578	3,860	16.3%
largo_real	benigno	2,266	486	486	3,238	13.7%
conversacional	benigno	1,692	363	362	2,417	10.2%
meta_texto	benigno	677	145	144	966	4.1%
documentacion	benigno	630	136	135	901	3.8%
Total		16,888	3,301	3,510	23,699	100%

Family glossary: plain = direct natural-language injection/jailbreak prompts; red_team = multi-turn manipulation and indirect injection through tool or document context; paraphrase = mechanically derived rewordings of attack seeds; multilingual_attack = machine-translated attack seeds; corto = short real user-style requests; largo_real = continuous long-form prose; conversacional = multi-turn human-assistant dialogue; meta_texto = security-themed text about attacks without live payloads; documentacion = technical documentation excerpts.

Language coverage:

Language(s)	Rows	Where
en	10,996	Verified-English rows across the five benign families
(unmarked, English-derived)	10,505	Attack families plain, red_team, paraphrase
es, fr, de, pt, it, nl, pl, ru	1,812	multilingual_attack translations (fr 228, es 227, pt 227, it 227, nl 227, pl 227, de 225, ru 224)
~40 additional locales	386	Short benign prompts from MASSIVE, native locale marked

No single multilingual locale exceeds 12.6% of the multilingual bucket.

Split protocol

Splits were produced by a group-aware MinHash near-duplicate pipeline (seed 42):

1. Texts are shingled into character 8-grams and embedded into 128-permutation MinHash signatures.
2. Locality-sensitive hashing (32 bands x 4 rows) with a Jaccard threshold of 0.5 groups exact and near-duplicate texts into groups.
3. Groups - not individual rows - are assigned to train / validation / test at 70/15/15, so near-duplicates can never straddle splits.
4. Each machine translation is force-paired with its seed before assignment (cross-language pairs cannot be linked by n-gram similarity), guaranteeing co-splitting of seed and translation.
5. Post-assignment audit: 0 violations (no exact or near-duplicate pair crosses a split), split ratios within +/-10 percentage points of target.

Data fields

Column	Type	Description
text	string	Document or prompt content
label	string	ataque (attack) or benigno (benign)
family	string	One of the nine families listed above
source	string	Source identifier (see Licensing)
license	string	Upstream license for the row's origin material
lang	string	Language tag; empty for the English-derived attack families
char_len	int	Character length of text
provenance	string	Row lineage: source item reference; for derived rows, the seed identifier plus transform details; for translated rows, seed id, locale, and model
split	string	Present only in pool_v15.parquet and corpus_v15.csv

Files: data/train.parquet, data/validation.parquet, data/test.parquet (standard split naming), plus the complete pool as data/pool_v15.parquet and data/corpus_v15.csv.

Provenance and licensing

Every row retains its upstream license. Sources, verified against their official repositories or dataset metadata, and row counts:

Source	Origin	License	Rows
hh_rlhf	Anthropic hh-rlhf	MIT	6,437
injecagent	InjecAgent benchmark	MIT	2,944
aegis_2.0	NVIDIA Aegis AI Content Safety 2.0	CC-BY-4.0	2,541
dolly_15k	Databricks Dolly-15k	CC-BY-SA-3.0	2,209
wikipedia_en	Wikimedia Wikipedia (en)	CC-BY-SA-3.0+GFDL	1,551
c4_en	C4 (domain-filtered)	ODC-BY	1,541
python_docs	Python official documentation	PSF	1,529
gsm8k	OpenAI GSM8K	MIT	1,354
advbench	AdvBench behaviors	MIT	742
oasst1	OpenAssistant oasst1	Apache-2.0	872
prompt_injections	deepset prompt-injections	Apache-2.0	394
massive	Amazon MASSIVE	CC-BY-4.0	390
rfc_editor	RFC Editor	IETF-Trust-Legal-Provisions	273
jbb_behaviors	JailbreakBench JBB-Behaviors	MIT	154
helpsteer2	NVIDIA HelpSteer2	CC-BY-4.0	479
gutenberg	Project Gutenberg	Public-domain-US	146
agentdojo	ETH Zurich AgentDojo	MIT	112
bipia	Microsoft BIPIA	MIT	31

Notes:

• The corpus-level license is CC-BY-4.0. Where an upstream license carries attribution or share-alike terms (for example CC-BY-SA on Dolly and Wikipedia, ODC-BY on C4), those upstream terms continue to govern the underlying text.
• The rfc_editor contributions reproduce text under the IETF-Trust-Legal-Provisions license, which permits reproduction with attribution. gutenberg rows are US public-domain literature (Public-domain-US); wikipedia rows carry the dual CC-BY-SA-3.0+GFDL grant.
• Restriction notice: an additional 1,416 rows derived from the proprietary zn mutation set (license value Propietario-interno-zn(sin redistribucion), redistribution never granted) are not included in this release; they are reserved privately and do not appear in any released file.
• No source with unknown or non-permissive licensing was used. Sources rejected during clearance (non-commercial or unlicensed user-generated dumps) contribute zero rows.

Multilingual translation methodology

The multilingual_attack family contains 1,812 machine-translated attack texts in eight locales (es, fr, de, pt, it, nl, pl, ru):

• Seeds were selected deterministically by stable hash from existing labeled attacks in the pool (no ad-hoc picking, reproducible selection).
• Translation used the Qwen3-32B model served through Amazon Bedrock (qwen.qwen3-32b-v1:0) with temperature 0.1 and extended thinking disabled.
• Model responses were frozen to disk per seed before any downstream use, so rebuilding the dataset replays identical payloads instead of re-invoking the model.
• Automated validation kept only strict outputs: a complete JSON envelope covering all eight locales, output differing from its seed, uniqueness across locales, and script-direction checks (Cyrillic-majority for Russian, Latin-script elsewhere).
• Total translation spend was USD 0.07 for the full bucket, recorded per-call in a cost ledger.

These translations are machine-generated: quality is medium, occasional mechanical errors and unnatural phrasing remain, and no professional human translation review has been performed. Treat the multilingual bucket as a cross-lingual robustness surface, not as gold-standard natural text.

Limitations

• Translation quality is medium. Machine-generated translations may contain errors; sample and review them before high-stakes use.
• Declared composition gaps. Relative to the design blueprint, three families finished under target: plain at 61% (4,265 of 7,000), documentacion at 45% (901 of 2,000), and long-form prose at 81% (3,238 of 4,000). The gaps come from gated upstream sources and finite document pools; they were left open rather than silently patched.
• Paraphrases are mechanical. The paraphrase family uses a fixed synonym table, safe clause reordering, and fixed contextual prefixes, so it anchors surface-form variation but does not provide deep semantic diversity.
• Legacy evaluation overlap. 135 rows have exact twins in an older evaluation subset; if that subset is reused for external evaluation, exclude its twins from one side first.
• Adversarial content. The corpus intentionally contains jailbreaks, injection payloads, and manipulative text. Use it for defensive research and evaluation of detection systems under applicable policies.
• An additional 1,416 held-out rows (synthetic mutants derived from our internal production corpus) are reserved privately and are not part of this public release.

Citation

@dataset{zn_prompt_injection_corpus_v15,
  title     = {ZN Prompt-Injection Evaluation Corpus (v15)},
  author    = {ZN},
  year      = {2026},
  version   = {v15},
  url       = {https://github.com/tljohnsilver/zn}
}

• Website: https://usezn.com
• Repository: https://github.com/tljohnsilver/zn