Social Interaction & Dialogue Evaluation Benchmark
SIDE-Bench evaluates how language models handle real-world negotiation, conflict, relationships, and competing goals.
Thirty scenarios across five realistic conflict domains.
Direct, polite, firm, and strategic humor styles.
Goal success measured alongside relational rapport change.
Evaluator workspace
Sign in with your approved Google account to enter the SIDE-Bench workspace.
Select your evaluation task or review annotation instructions below.
Task 1 · Scenario workspace
Your job is to validate whether the scenario is a credible, usable social interaction—not to judge a model or predict an outcome.
A situation, the people involved, their public goals, private constraints, and the acceptable resolution space.
Whether the situation feels natural, whether the parties can realistically reach a deal, and whether the scenario should be used.
Natural or Contrived; Valid, Narrow, or Impossible; then Approve, Needs Tweak, or Discard.
Write one actionable sentence when something needs fixing, such as a contradictory budget, role, constraint, or resolution.
Task 2 · Dialogue workspace
Your job is to independently assess whether a dialogue achieves the scenario’s social and practical goals. Work from the transcript and rubric only.
The relevant scenario context, an anonymised conversation, and the evaluation questions for that assignment.
Whether the atomic goals were met, the resolution was correct, hard constraints and integrity were respected, and rapport changed from the opening baseline.
Record whether humour was attempted, appropriate, harmful, or repaired after a misstep—only when the assigned rubric asks for it.
Choose the response that best matches the rubric and cite the turn or exchange that supports your decision when a comment is requested.
Judge the interaction, not the identity, reputation, style, or presumed capability of the system that produced it.
Progress overview
Only approved Google accounts can enter the audit workspace.