Stanford UniversityAI Measurement ScienceAIMS — AI Measurement Science

Medicine

Thousand Voices of Trauma

Thousand Voices of Trauma: 3,000 synthetic Prolonged-Exposure therapy conversations (500 clients x 6 phases, generated by Claude Sonnet 3.5) plus an emotional-trajectory benchmark in which 8 LLMs rate the emotional intensity (1-10) of each PE-therapy phase per client, scored against the Claude Sonnet 3.5 baseline trajectory.

3,000items
8subjects
CC-BY-NC-4.0license
medicinedomain
textmodality
item-level responses released
Saturation status: Unknown
Original sourcePaper← All benchmarks

Response matrix

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

Thousand Voices of Trauma response matrix: AI models (rows) against items (columns)
lowhighUnobserved

Scale: emotional intensity {1, ..., 10}

Subjects

  1. 1Mistral Small6.866
  2. 2Mistral 7B Instruct6.5193
  3. 3Llama 3 70B Instruct6.4461
  4. 4Mistral Large6.3487
  5. 5Llama 3.1 70B Instruct6.269
  6. 6Amazon Nova Pro5.9713
  7. 7Llama 3.1 8B Instruct5.954
  8. 8Llama 3 8B Instruct5.729
Full data on Hugging FaceBack to the gallery