guard-glp-benign / README.md
ddidacus's picture
Upload README.md with huggingface_hub
fc96e9b verified
|
Raw
History Blame Contribute Delete
689 Bytes
metadata
license: mit

Guard-GLP Benign Conversations

Sanitized collection of benign multi-turn conversations, using train splits only.

Composition

Source Samples
lmsys 942,157
wildchat 524,274
wildguard 7,778
wildjailbreak 128,781
Total 1,602,990

Total tokens: 1,124,232,118 (Llama 3.2 1B Instruct tokenizer)

Sanitization

  • LMSYS Chat 1M & WildChat: filtered via OpenAI moderation labels (all flagged categories removed)
  • WildGuardMix: removed adversarial and harmful-labeled prompts/responses
  • WildJailbreak: removed entries with harmful data types

Only benign conversations are retained.