ArnoldMoya/execoconut-dataset

# execoconut-dataset

ExecCoCoNuT (Execution CoCoNuT): Python code execution traces for continuous latent thought training.

## Dataset Description

This dataset contains 1,000+ Python code snippets with execution traces, designed for training language models to reason about program state in continuous latent space.

## Dataset Statistics

- **Samples**: ~1,000 (train: 900, val: 50, test: 50)
- **Variable scope**: 6 integer variables (a-f)
- **Operations**: +, -, *, // (integer division)
- **Control flow**: Sequential only (no loops/branches)
- **Value range**: [-100, 100]
- **Snippet length**: 3-8 instructions per sample

## Data Format

Each sample is a JSON object with three fields:

```json
{
"question": "a = 3\nb = a + 2\nc = b * a",
"steps": [
"State: {\"a\": 3}",
"State: {\"a\": 3, \"b\": 5}",
"State: {\"a\": 3, \"b\": 5, \"c\": 15}"
],
"answer": "c = 15"
}
```

- **question**: Multi-line Python code snippet
- **steps**: Execution trace showing state after each instruction
- **answer**: Final variable assignment (the target prediction)

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("ArnoldMoya/execoconut-dataset")

# Access splits
train = dataset['train']
val = dataset['validation']
test = dataset['test']

# Example
for sample in train.take(1):
print(sample['question'])
print(sample['steps'])
print(sample['answer'])
```

## Splits

- `train.jsonl`: Training split (90%)
- `validation.jsonl`: Validation split (5%)
- `test.jsonl`: Test split (5%)

## Paper

Introduced in "ExecCoCoNuT: Latent Code Execution via Continuous Thought Chains" (2026)

Built with [COCONUT](https://github.com/facebookresearch/coconut) (Meta FAIR, 2024)

## License

CC0 1.0 Universal