RUBYGLASK/llm-redteam-corpus-taxonomy-extended

LLM Red-Team Corpus + Taxonomy — Extended (Non-Commercial)

⚠️ Non-commercial use only. For commercially-usable content, see rubyglask/llm-redteam-corpus-taxonomy-core — CC-BY-4.0.

Status: Phase 1 (2026-Q3) — first data upload targets Phase 1 Week 5. This card is a skeleton awaiting the initial data release.

Dataset Description

The extended tier of the LLM Red-Team Corpus + Taxonomy project. Aggregates upstream sources under non-commercial licenses (CC-BY-NC-4.0, CC-BY-NC-SA-4.0). Kept separate from core so that core remains commercially usable while researchers still have access to a broader corpus for non-commercial red-team / safety work.

Included upstream (~34K target)

• BeaverTails (PKU-Alignment) — CC-BY-NC-4.0
• ALERT (Babelscape) — CC-BY-NC-SA-4.0
• LMSYS Toxic-Chat — CC-BY-NC-4.0
• Anthropic Persuasion — CC-BY-NC-4.0

Data Structure

Same canonical schema as core (see core dataset card). license_tier is fixed at extended for all rows in this dataset.

How to load

from datasets import load_dataset
ds = load_dataset("rubyglask/llm-redteam-corpus-taxonomy-extended")

License

Each upstream source retains its own non-commercial license. This aggregation is distributed under a mixed non-commercial umbrella — users must comply with each source's individual license.

Acceptable uses:

• Academic research
• Non-commercial safety evaluation
• Defensive red-team research at non-profit organizations

Not acceptable:

• Commercial product integration
• Paid guardrail / evaluation services
• Any use that would violate the source-level NC constraint

Code repository

Build scripts, changelog, methodology: https://github.com/Rubyglask/llm-redteam-corpus-taxonomy