Jarrodbarnes/donnabench-tasks

Personal assistant tasks from observed conversations

Benchmark task candidates for AI personal assistants, inferred by reading conversations rather than written by hand. A simulated person with a rich persona (a retired teacher, a tiling contractor, a part-time physiotherapist with two children, a consultant, a student) talked to an assistant that could only act through tools over a closed world: calendar, inbox, contacts, a restaurant guide, a store, a marketplace with buyers who answer, and subscriptions. Every conversation was recorded in full and reviewed.

• 25 conversations, 129 turns in the traces config.
• 17 tasks in the tasks config, each with the evidence it came from.
• Capabilities targeted: personalization, clarification, permission, recovery, proactivity.
• Assistant model: z-ai/glm-5.3-flash; person model: z-ai/glm-5.3-flash; reviewing model: gpt-5.6-sol.

How the tasks were made

A reviewing model read each conversation in full: the person's goals and what they privately knew, every message, every tool call with its arguments and result, the person's private thoughts and their final verdict. It wrote down what the assistant did well and badly against the five capabilities, with a verbatim quote for each observation. A second pass read all observations together and proposed tasks. A third pass re-read only the cited turns, one fresh call per citation, and confirmed or refuted it; a task keeps only confirmed evidence and was dropped without any. Of 20 proposed tasks, 17 survived. No pattern matching or hand-written rules took part.

tasks

field	meaning
task_id	task-001 …
title	what the task tests, in one line
capability	one of personalization, clarification, permission, recovery, proactivity
persona	which person the situation belongs to
situation	what the world must contain and what the person is trying to do
opening_message	the message the person opens with, in their voice
expected_behavior	what a competent human assistant does
check_on	state, reply or both: where a script checks the outcome
check_positive	conditions that must hold after the conversation
check_negative	conditions that must not: nothing else changed, nothing sent that should not be
evidence	the observed turns this was inferred from: episode, turn, quote, and the verifier's reason
recurrence	how many conversations showed it
why_it_matters	one sentence on the cost to the person

traces

field	meaning
episode_id, turn	which conversation and which exchange
persona, person_name, goals	who the person is and what they wanted this session
user_message	what the person typed
private_thought	what the person was thinking; the assistant never saw it
tool_calls	JSON list of {name, args, result, is_error}, in order
assistant_reply	what the assistant said back
agent_error	set when the assistant's turn failed
end_reason, satisfaction, verdict	how the conversation ended and what the person concluded, 1 to 5
cited_by	task ids whose evidence points at this turn

Loading

from datasets import load_dataset
tasks = load_dataset("Jarrodbarnes/donnabench-tasks", "tasks", split="train")
traces = load_dataset("Jarrodbarnes/donnabench-tasks", "traces", split="train")

The worlds, personas, simulator and review code live in the observe/ directory of the DonnaBench repository.