keenanpepper/textworld-state-tracking

TextWorld State Tracking Dataset

Dataset Description

This dataset contains 27,145 examples for evaluating language models' ability to track object states across narrative contexts. The data is derived from TextWorld, an interactive text-based game environment, and formatted for state tracking evaluation.

Dataset Summary

Each example presents a narrative context (a sequence of game observations and actions) along with:

• An entity (object) being tracked
• Its positive_state (the true current state)
• A negative_state (a mutually exclusive contrastive state)

The task is to determine which state correctly describes the entity given the context.

Key Statistics:

• Total examples: 27,145
  • Train: 24,076 examples
  • Dev: 3,069 examples
• Tracked properties: 3 (open, closed, eaten)
• Average context length: ~2,850 words (median: 2,064 words)
• Context length range: 53 - 12,855 words (25th-75th percentile: 950 - 4,180 words)
• Unique entities: 13 distinct objects tracked
• Entities tracked: Various objects from TextWorld environments (doors, containers, food items, etc.)

Supported Tasks

1. Binary State Classification: Given context and entity, predict which of two states is correct
2. State Tracking: Track state changes of entities across narrative sequences
3. Reading Comprehension: Understanding implicit and explicit state information from text

Languages

The dataset is in English.

Dataset Structure

Data Instances

Each instance contains:

{
  "context": "Text describing the game state and actions taken",
  "entity": "Name of the object being tracked",
  "positive_state": "The correct current state (e.g., 'closed', 'eaten', 'open')",
  "negative_state": "A contrastive incorrect state (e.g., 'open', 'not eaten', 'closed')"
}

Example:

{
  "context": "-= Bedroom =-\nYou see a closed chest drawer. You see an antique trunk...\n> open chest drawer\nYou open the chest drawer.",
  "entity": "chest drawer",
  "positive_state": "open",
  "negative_state": "closed"
}

Data Fields

• context (string): Progressive narrative showing the game state, observations, and actions taken. Includes room descriptions and action-response pairs.
• entity (string): The specific object whose state is being tracked (e.g., "wooden door", "apple", "refrigerator")
• positive_state (string): The ground truth current state of the entity. One of: "open", "closed", or "eaten"
• negative_state (string): A mutually exclusive state that does NOT describe the entity. One of: "closed", "open", or "not eaten"

Data Splits

	Train	Dev
Examples	24,076	3,069

The splits are based on distinct TextWorld game traces, ensuring no overlap in source trajectories.

Dataset Creation

Source Data

The dataset is derived from TextWorld game traces generated using the TextWorld framework. TextWorld creates procedurally generated text-based games with consistent world states.

Initial Data Collection

1. Game Generation: TextWorld environments were generated with various objects and properties
2. Trace Collection: Agent trajectories were collected, capturing sequences of actions and observations
3. State Extraction: Ground truth entity states were extracted from the game engine's internal state

Data Processing

The raw TextWorld traces underwent several processing steps:

1. State Annotation: Each game step was annotated with the current states of all entities

2. Context Assembly: Progressive contexts were built by accumulating observations and actions

3. Quality Filtering:

   • Minimum context length: 50 words (ensures non-trivial examples)
   • Entity mention verification: Entity must be mentioned in context
   • Property mention verification: State must be mentioned in context
   • State balance filtering: Entities must appear with multiple states (prevents trivial learning)

4. Contrastive Pair Creation: For each entity-state pair, a mutually exclusive contrastive state was assigned

Annotations

Annotation process

Annotations are automatically extracted from TextWorld's game engine, which maintains perfect ground truth about entity states. No human annotation was required.

Who are the annotators?

N/A - Automatically generated from game engine state.

Personal and Sensitive Information

The dataset contains only synthetic text from procedurally generated games. No personal or sensitive information is present.

Considerations for Using the Data

Social Impact of Dataset

This is a synthetic dataset for research purposes, with no direct social impact concerns.

Discussion of Biases

The dataset reflects the structure and vocabulary of TextWorld environments:

• Limited to household/indoor settings
• Stereotypical room layouts and object placements
• Formulaic language patterns from the text generation templates

Other Known Limitations

1. Limited Property Types: Only tracks 3 property types (open/closed/eaten)
2. Limited Entity Diversity: Only 13 unique entities across all 27K examples
3. Synthetic Language: Generated text has repetitive patterns and is not representative of natural language
4. Domain-Specific: Focused on household objects and simple spatial relations
5. State Simplicity: Real-world states are more complex and nuanced
6. Long Contexts: Average of ~2,850 words may exceed context windows of smaller models
7. Context Redundancy: Progressive contexts include many repeated observations and may allow shortcuts without full state tracking

Additional Information

Dataset Curators

Created as part of research on language model state tracking capabilities.

Licensing Information

MIT License

Citation Information

@dataset{textworld_state_tracking_2024,
  title={TextWorld State Tracking Dataset},
  author={Anonymous},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/YOUR_USERNAME/textworld-state-tracking}}
}

Contributions

Built using the TextWorld framework by Microsoft Research.