Knowledge

WikiContradict

WikiContradict: real-world knowledge-conflict QA mined from Wikipedia. Each item is a question whose answer is contradicted across two Wikipedia passages; LLM answers are graded by humans as Correct / Partially correct / Incorrect under several prompt (RAG) templates.

48items
5subjects
Apache-2.0license
knowledgedomain
textmodality
item-level responses released
Saturation status: Unknown

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

WikiContradict response matrix: AI models (rows) against items (columns)
lowhighUnobserved

Scale: {0, 0.5, 1} (Incorrect / Partially correct / Correct)

Subjects

  1. 1llama-3-70b-instruct0.6542
  2. 2mixtral-8x7b-instruct-v0-10.6229
  3. 3gpt-40.6083
  4. 4llama-2-70b-chat0.5917
  5. 5mistral-7b-instruct-v0-20.5813