---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- synthetic
- agent
- alfworld
- household
pretty_name: "Interactive Household Task Trajectory Dataset (Teacher: Qwen/Qwen2.5-72B-Instruct-AWQ, vLLM AWQ)"
---

# Interactive Household Task Trajectory Dataset

This dataset contains **synthetic multi-step interactive task trajectories**
generated from a fully self-constructed environment simulator.

The data follows the same multi-turn `messages` structure as the
official training dataset provided by the competition organizers.

## Format
- `messages`: list of objects with:
  - `role`: system | user | assistant
  - `content`: string
- Assistant turns follow:
  - `Think: ...`
  - `Act: <action>`
- Environment feedback is provided as subsequent `user` messages.

## Environment & Data Generation
- All environments, tasks, and trajectories are generated from scratch.
- No external AgentBench or ALFWorld datasets were used.
- No data derived from external benchmarks was included.

## Teacher LLM
- Model: `Qwen/Qwen2.5-72B-Instruct-AWQ`
- Runtime: `vLLM` (AWQ)

## Generation notes
- Actions are planned by a deterministic planner to ensure valid transitions.
- A small fraction of recovery turns ("Nothing happens.") may be injected to improve robustness.

## Columns
- `messages`
- `task_type`
- `teacher_repo`
- `n_steps`
- `mistakes`