Open datasets.
Published, not described.

Work anyone can download, inspect and disagree with. Each release lists what it contains, how it was built, and where it falls short — the last part being the one that makes the rest worth reading.

Published datasets

01 · 2026-08Published

empathy-conversations

Multi-turn dialogue centred on emotional support — parenting pressure, mental-health concerns, and the ordinary conversations where tone matters more than information.

Released as our first open artefact. The intent is narrow and deliberate: most conversational corpora optimise for task completion, so models trained on them learn to resolve rather than to sit with a problem. This set is the opposite case.

Rows4,010
LanguageEnglish
FormatJSON
Turns~10 per conversation
TasksDialogue modelling · generation
LicenseSee dataset card

Limitations

  • English only. It does not yet cover the Indic and code-switch settings our production systems run in.
  • Emotional-support scenarios are a narrow slice of conversation; do not treat coverage here as coverage of dialogue generally.
  • Not a clinical resource, and not validated for any diagnostic or therapeutic use.

Mirrors  ·  6