Wikibench

HellaSwag

reasoning

A challenging commonsense natural language inference dataset that uses Adversarial Filtering to create questions trivial for humans (>95% accuracy) but difficult for state-of-the-art models, requiring completion of sentence endings based on physical situations and everyday activities

Leaderboard

Showing 20 of 27 results

  1. Claude 3 Opus
    95.4%
    i
  2. GPT-4
    95.3%
    i
  3. Gemini 1.5 Pro
    93.3%
    i
  4. MiMo-V2.5-Pro
    89.8%
    i
  5. Claude 3 Sonnet
    89.0%
    i
  6. Command R+
    88.6%
    i
  7. Hermes 3 70B
    88.2%
    i
  8. Qwen2 72B Instruct
    87.6%
    i
  9. Gemini 1.5 Flash
    86.5%
    i
  10. Gemma 2 27B
    86.4%
    i
  11. Claude 3 Haiku
    85.9%
    i
  12. Llama 3.1 Nemotron 70B Instruct
    85.6%
    i
  13. Qwen2.5 32B Instruct
    85.2%
    i
  14. Phi-3.5-MoE-instruct
    83.8%
    i
  15. Mistral NeMo Instruct
    83.5%
    i
  16. Qwen2.5-Coder 32B Instruct
    83.0%
    i
  17. Gemma 2 9B
    81.9%
    i
  18. Granite 3.3 8B Base
    80.1%
    i
  19. Gemma 3n E4B
    78.6%
    i
  20. Gemma 3n E4B Instructed LiteRT Preview
    78.6%
    i