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:

{
  "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

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 (Meta FAIR, 2024)

License

CC0 1.0 Universal