Tralalabs/simplemath-12k

Tiny Addition Dataset for 1M Models

BROOOOOOOOOOOOOO 💀💀💀💀

A tiny synthetic arithmetic dataset designed for training or fine-tuning very small language models around 1M parameters on basic addition patterns.

The dataset focuses on simple additions such as:

1+1=2
5+5=10
12+8=20

It is intentionally narrow, repetitive, and clean so that small models can learn the task without needing a large vocabulary or complex reasoning corpus.

Dataset Summary

This dataset contains short instruction-response examples for basic addition from 0 to 20.

It includes multiple prompt styles:

• direct equation format
• question-answer format
• instruction format
• short explanation format

Example:

{"text":"### Instruction:\nWhat is 5+5?\n\n### Response:\n10"}

Files

File	Description
train.jsonl	Training split with 12,000 examples
validation.jsonl	Validation split with 1,200 examples
README.md	Dataset card and documentation

Dataset Structure

Each row is a JSON object with a single text field.

{
  "text": "### Instruction:\n1+1=\n\n### Response:\n2"
}

Example Rows

{"text":"### Instruction:\n1+1=\n\n### Response:\n2"}
{"text":"### Instruction:\nWhat is 5+5?\n\n### Response:\n10"}
{"text":"### Instruction:\nExplain 2+3.\n\n### Response:\n2+3 means start with 2 and add 3. The answer is 5."}

Intended Use

This dataset is intended for:

• toy language model training
• tiny transformer experiments
• testing arithmetic pattern learning
• quick fine-tuning experiments
• educational ML demos
• small-scale instruction tuning

Recommended model scale:

Model Size	Expected Result
~1M parameters	Can learn narrow addition patterns
~10M parameters	More stable and less fragile
~30M parameters	Better generalization on simple arithmetic formats

Recommended Training Setup

For a very small model, keep the setup simple:

context_length: 64 or 128
vocab_size: small tokenizer preferred
task: causal language modeling
epochs: 3 to 10
learning_rate: 1e-4 to 5e-4
batch_size: as large as your GPU/CPU allows

A 1M parameter model should train on this dataset as a narrow pattern-learning task, not as a general math reasoning model.

Limitations

This dataset is extremely small and narrow.

It does not teach:

• advanced arithmetic
• subtraction
• multiplication
• division
• multi-step reasoning
• word problems
• robust out-of-distribution math

The model may memorize patterns instead of learning true arithmetic. That is expected for tiny models.

Biases and Risks

The dataset is synthetic and contains only basic arithmetic examples. It has very low social or demographic bias risk, but it can create overconfidence if used to claim that a model has broad mathematical reasoning ability.

Data Generation

The dataset was generated synthetically using Python.

Numbers were sampled from the range:

0 to 20

Prompt styles were randomly mixed to create format variety while keeping the task simple.

License

This dataset is released under the Creative Commons Zero v1.0 Universal license.

You may use, copy, modify, distribute, and train models on this dataset without asking permission.

License keyword:

license: cc0-1.0

Citation

If you use this dataset, citation is optional because it is released under CC0.

Optional citation format:

@dataset{tiny_addition_1m,
  title = {Tiny Addition Dataset for 1M Models},
  author = {Synthetic Dataset},
  year = {2024},
  license = {CC0-1.0}
}

Quick Loading Example

from datasets import load_dataset

dataset = load_dataset("json", data_files={
    "train": "train.jsonl",
    "validation": "validation.jsonl"
})

print(dataset["train"][0]["text"])

Suggested Prompt Format

### Instruction:
What is 5+5?

### Response:
10

Final Note

This dataset is intentionally simple. For a tiny model, simple is powerful.