cmu-lti/stateful

---
language:
- en
tags:
- software-engineering
- code
- swe-bench
- stateful
- user-modeling
- theory-of-mind
size_categories:
- 1K<n<10K
task_categories:
- text-generation
- question-answering
---

# Stateful SWE Dataset

## Dataset Summary

The **Stateful SWE Dataset** extends the [cmu-lti/interactive-swe](https://huggingface.co/datasets/cmu-lti/interactive-swe) dataset with user profile assignments for studying stateful interactions in software engineering tasks. Each instance from the original interactive-swe dataset is enriched with a randomly assigned user profile that defines interaction preferences and coding standards.

This dataset enables research into:
- **Theory of Mind (ToM)** modeling for AI agents
- **Stateful user interactions** in software engineering
- **Personalized code assistance** based on user preferences
- **User behavior modeling** in programming contexts

## Dataset Details

- **Total instances**: 500
- **User profiles**: 15 distinct profiles
- **Base dataset**: cmu-lti/interactive-swe
- **Assignment**: Random profile assignment with seed 42
- **Version**: 1.0.0

## Dataset Structure

### Original Interactive-SWE Columns (15 columns)
All original columns from cmu-lti/interactive-swe are preserved:
- `repo`: Repository name
- `instance_id`: Unique identifier from original dataset
- `base_commit`: Base commit hash
- `patch`: Code changes
- `test_patch`: Test-related changes
- `problem_statement`: Description of the issue
- `hints_text`: Additional hints
- `created_at`: Original timestamp
- `version`: Version information
- `FAIL_TO_PASS`: Test information
- `PASS_TO_PASS`: Test information
- `environment_setup_commit`: Environment setup details
- `difficulty`: Problem difficulty level
- `original_issue`: Link to original issue
- `files`: Comma-separated list of affected files

### New Stateful Columns (8 columns)
- `user_profile_id`: Assigned user profile identifier
- `user_roleplay_prompt`: Second-person narrative describing the user
- `interaction_preferences`: JSON string with verbosity, timing, and response style preferences
- `coding_preferences`: Comma-separated string of user's technical preferences
- `stateful_instance_id`: New unique identifier for stateful instances
- `assignment_seed`: Random seed used for profile assignment
- `dataset_version`: Version of the stateful dataset
- `created_at_stateful`: Timestamp when stateful instance was created

## User Profile Types

The dataset includes 15 diverse user profiles with varying:

- **Verbosity preferences**: concise vs verbose
- **Question timing**: upfront vs ongoing clarification
- **Response style**: short vs long responses
- **Coding preferences**: frameworks, testing, documentation, etc.

## Usage Example

```python
from datasets import load_dataset
import json

# Load the stateful dataset
dataset = load_dataset("cmu-lti/stateful", split="test")

# Access an instance with its user profile
instance = dataset[0]
print(f"Problem: {instance['problem_statement']}")
print(f"User Profile: {instance['user_profile_id']}")

# Parse interaction preferences from JSON string
prefs = json.loads(instance['interaction_preferences'])
print(f"Interaction Style: {prefs}")

# Parse coding preferences from comma-separated string
coding_prefs = instance['coding_preferences'].split(',')
print(f"Coding Preferences: {coding_prefs[:3]}...") # Show first 3
```

## Citation

If you use this dataset, please cite both the original interactive-swe dataset and this stateful extension:

```bibtex
@dataset{stateful_swe_2025,
title={Stateful SWE Dataset: User Profile Extensions for Interactive Software Engineering},
author={CMU ToM-SWE Team},
year={2025},
url={https://huggingface.co/datasets/cmu-lti/stateful}
}
```

## License

This dataset follows the same license as the original cmu-lti/interactive-swe dataset.

## Dataset Creation

Created using the ToM-SWE framework for Theory of Mind modeling in software engineering contexts.