Sakai0920/alfworld-distilled-qwen3-32b

ALFWorld Distilled Trajectories (Model-Generated)

Overview

This dataset contains interaction trajectories collected by allowing a teacher language model to interact with the ALFWorld environment using a ReAct-style policy.

These trajectories are intended for use as model-generated supervision signals in agent training via supervised fine-tuning (SFT).

Data Generation Process

Trajectories were generated by running the ALFWorld simulator and allowing a teacher model (Qwen3-32B-AWQ) to solve tasks through iterative reasoning and action selection.

All Thought and Action steps contained in this dataset are generated by the teacher model itself during environment interaction.

No ground-truth or oracle trajectories provided by ALFWorld were used.

The ALFWorld environment was used solely as an interactive simulator during data generation and not as a source of supervised labels or demonstration data.

All supervision signals in this dataset originate from model-generated outputs.

Compliance Note

We acknowledge that ALFWorld's official train split constitutes benchmark data.

However, this dataset does not contain any ground-truth solutions, scripted policies, or human demonstrations from the benchmark.

Instead, it consists entirely of trajectories generated by a language model interacting with the environment.

We believe this setup complies with the competition rules, as the training data is derived from model outputs rather than benchmark-provided solutions.

Teacher Model

• —Qwen3-32B-AWQ

Intended Use

This dataset is intended for:

• —Agent policy distillation
• —Supervised fine-tuning (SFT) of language models for interactive environments

Format

Each sample contains:

• —messages: multi-turn interaction history