viviencheng/butler-bench-v2

Butler-Bench v2

A personal-assistant tool-use benchmark. An agent must retrieve a user's private digital artifacts, resolve stale vs. current evidence, infer hidden preferences, and execute the correct terminal tool call under a one-tool-call-per-turn protocol.

Structure

• 100 shared task templates × 5 personas = 500 episodes.
• For a given shared_task_id the instruction is identical across all personas; only the persona's artifacts and the (private) hidden rubric differ.
• 1436 artifacts across messages, emails, calendar, orders, notes, contacts, locations, memberships, files.

Files (this public repo)

• episodes.csv — one row per episode: the rendered agent prompt + tool metadata. No answers.
• shared_tasks.csv — the 100 task templates (family, domain, terminal tool, required fields, tools).
• artifacts.jsonl — the per-persona artifact worlds the agent retrieves from.
• personas.json — persona axes and stable preferences.

Difficulty tiers (in the private rubric)

stated (answer written directly) · drifted (old value superseded by newer) · derived (inferred from multiple artifacts) · open (no direct evidence; safe default, rarely scored).

⚠️ Rubrics are private (eval-only)

The hidden answer key (rubric.csv, with correct_value, per-slot tiers/evidence, and the full target_action_json) is NOT in this repo — it lives in a separate private eval repo. Public files deliberately contain no correct_value and no target_action_json.

Counts at packaging: 5 personas / 100 tasks / 500 episodes / 1436 artifacts / 2395 rubric rows (private).