---
pretty_name: Rules vs. Examples
language:
  - en
license: mit
size_categories:
  - 100K<n<1M
---

## Paper Summary

This dataset accompanies the paper *LLMs Learn Better In-Context from Rules than from Examples*. The paper studies in-context learning in large language models using a suite of programmatically generated tasks across games, arithmetic, and linguistic inference. The dataset includes five task families, Set Game, Tapatan, Operator Function, Noun Class Agreement, and Lexical Category Inference.

## Dataset Structure

Repository files:

```bash
.
├── dataset.zip
├── LICENSE
└── README.md
```

Unzipping `dataset.zip` creates the following `data/` directory:

```bash
data/
├── lexical_category_inference/
│   ├── easy/shared/
│   ├── medium/conjunctive/
│   ├── medium/disjunctive/
│   ├── hard/conjunctive/
│   └── hard/disjunctive/
├── noun_class_agreement/
│   ├── easy/
│   ├── medium/
│   └── hard/
├── operator_function/
│   ├── easy/
│   ├── medium/
│   └── hard/
├── set_game/
│   ├── easy/
│   ├── medium/
│   └── hard/
└── tapatan/
    ├── move_sequence/
    │   ├── easy/
    │   ├── medium/
    │   └── hard/
    └── final_board_state/
        ├── easy/
        ├── medium/
        └── hard/
```

Each leaf directory under `data/` contains a `train.jsonl` file and a `test.jsonl` file.

## Data Format

Each JSONL row contains a task item, metadata describing the condition, and the gold label. The common top-level fields are:

- `task`: task identifier
- `task_display_name`: task name
- `difficulty`: difficulty level
- `condition`: task-specific condition metadata
- `item`: task-specific input fields
- `label`: gold answer
- `split`: `train` or `test`
- `source`: provenance metadata

## License

This dataset is released under the MIT License.