tau2-infinity-wg

Airline customer-service tasks mined by world-gen adversarial failure-search — each task is an artifact on which a target model diverges from an oracle model. Companion to vibrantlabsai/tau2-infinity; consumed by the tau2_infinity_wg Prime Intellect RL environment.

Overview

Unlike tau2-infinity (adaptively generated toward a target difficulty band), tasks here are discovered by hypothesis-disproof: generate a hypothesized failure mode, confirm the oracle resolves it cleanly while the target fails, then keep only cases where both hold. This biases the set toward borderline/adversarial scenarios rather than uniformly hard ones.

Each task pins an initial airline world (users, flights, reservations) and a set of objectives the agent must satisfy while handling a call from an LLM-simulated customer with a fixed personality and some hidden information.

Property	Value
Number of tasks	10
Failure categories	booking (7), policy (3)
Objectives per task	3–4 (mean 3.6)
Max turns	10–20 (median 15)
Customer simulator	LLM-driven, per-task personality + hidden info
Source	world-gen adversarial hypothesis-disproof

Dataset Schema

Column	Type	Description
task_id	string	Task identifier (e.g. failure-0189abb3)
task_description	string	Natural-language scenario the customer is calling about
customer_personality	string	Freeform persona for the customer simulator
customer_hidden_info	JSON string	Facts the customer knows but may not disclose unprompted
max_turns	int	Episode turn budget
objectives	JSON string	List of {id, description, type, verification_context} — judged at end of episode
n_objectives	int	Objective count
world_state	JSON string	world_export.world — entities, relationships, timeline
tools	JSON string	Tool specifications available to the agent (identical across tasks)
failure_category	string	e.g. booking, policy
failure_notes	string	Human-readable failure-mode notes from the mining run
raw_artifact	JSON string	Full original world-gen artifact (superset of above)

Tasks

task_id	failure_category	n_objectives	max_turns
failure-0189abb3	booking	4	16
failure-04c75bdf	booking	4	15
failure-2782cbce	booking	4	12
failure-2c501ab3	policy	4	16
failure-46b8621c	booking	4	15
failure-46e0e882	booking	3	12
failure-5214cbeb	policy	3	10
failure-86b0108a	policy	3	12
failure-96b8509e	booking	4	20
failure-dde59d2e	booking	3	15

Usage

from datasets import load_dataset
import json

ds = load_dataset("vibrantlabsai/tau2-infinity-wg", split="test")

for task in ds:
    print(task["task_id"], task["failure_category"])
    objectives = json.loads(task["objectives"])
    world = json.loads(task["world_state"])

License

Apache 2.0