license: other
tags:
- safety
- guardrails
- jailbreak
- content-moderation
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: valid
path: data/valid-* - split: eval
path: data/eval-*
dataset_info:
features: - name: messages
list:- name: role
dtype: string - name: content
dtype: string
splits:
- name: role
- name: train
num_bytes: 146173759
num_examples: 239292 - name: valid
num_bytes: 18753914
num_examples: 29738 - name: eval
num_bytes: 17377621
num_examples: 29598
download_size: 181361758
dataset_size: 182305294
- split: train
lego17440/guardrail-input-slm-data
Private training data for an encoding-aware input-safety guardrail SLM. Schema 0.1.0.
Format (chat)
chat:{messages:[system,user,assistant]}, assistant = minified verdict{"decision":"allow|block","categories":[...]}— for LLM-SFT (Qwen / Llama-Guard style).clf:{text, target}— for an encoder classifier (DeBERTa/ModernBERT).
Splits
Group-split by prompt (all encoded variants of a prompt share a split) — eval is held-out prompts across all encodings, a real generalization test.
Composition
Real public jailbreak/content-safety data + encoding augmentation (real prompts re-encoded: base64/ciphers/scripts) + OCR/ASR extraction-noise (multimodal readiness) + curated seeds & injection payloads.
Licensing (commercial build)
Built from permissive, human-written sources only (MIT / Apache-2.0 / CC-BY-SA-3.0 / ODC-BY): oasst1, dolly, Salad-Data, Anthropic/hh-rlhf red-team, in-the-wild-jailbreak, gandalf, JBB-Behaviors, jackhhao, deepset. Non-commercial sources (alpaca, no_robots, BeaverTails, toxic-chat = CC-BY-NC) and GPT-generated sets are excluded. Verify each source's terms for your jurisdiction before shipping.