Datasets:

beyoru
/

Qwen3.5-SafeProbe-data

license: cc-by-sa-4.0
language:
  - vi
  - en
task_categories:
  - text-classification
tags:
  - safety
  - guardrail
  - content-moderation
  - prompt-injection
  - jailbreak
pretty_name: Qwen3.5-SafeProbe training data
extra_gated_prompt: >-
  Access is granted manually for research on AI safety only. This dataset
  contains harmful, offensive and sexual text.
extra_gated_fields:
  Affiliation: text
  Intended use: text

Qwen3.5-SafeProbe — training data

Training and held-out test data for beyoru/Qwen3.5-SafeProbe-2B.

⚠️ Contains harmful, hateful, sexual and otherwise offensive text, collected for training safety classifiers.

File	Rows	unsafe	safe
`misa/train.jsonl`	31,832	11,117	20,715
`misa/test.jsonl`	3,565	1,315	2,250
`standard/train.jsonl`	71,774	29,257	42,517
`standard/test.jsonl`	7,908	3,287	4,621
`misa/token_head_train.jsonl`	12,000	span-labelled long inputs + 3,500 long agent-internal prompts (safe, hard negatives)
`standard/token_head_train.jsonl`	8,500	span-labelled long inputs

misa/: MISA policy (Vietnamese law + enterprise-agent rules). standard/: public-benchmark convention.
Row format: {"id", "source", "label": "safe"|"unsafe", "messages": [OpenAI-style]}; token-head rows add unsafe_span / safe_span (the inserted request inside a long input).
Test splits are held out by scenario; training data was decontaminated against WildGuardTest, ToxicChat, OpenAI Moderation, Aegis 2.0, JBB, SimpleSafetyTests, XSTest and MultiJail.

Includes material derived from WildGuardMix (ODC-BY), Aegis 2.0 / Nemotron-Safety-Guard (CC-BY-4.0), Nailyk14/prompt-safety-multilingual-v2 (mixed, incl. CC-BY-SA-4.0) and WildChat (ODC-BY), plus synthetic and proprietary data.