| --- |
| language: [en] |
| tags: [arithmetic, pretraining, sub-10M, ArithMark] |
| --- |
| # Arithmetic-XL |
|
|
| ## 12 Million Verified Arithmetic Examples for Language Model Pretraining |
|
|
| > *"Arithmetic is one of those things that humans learn in primary |
| > school but language models somehow manage to forget halfway through |
| > training."* |
|
|
| Arithmetic-XL is a **large-scale arithmetic pretraining dataset** |
| containing **over 12 million procedurally generated and verified |
| examples** designed to improve numerical reasoning in language models. |
|
|
| The dataset was created with one objective: |
|
|
| **Teach models how to think about arithmetic instead of simply |
| memorizing answers.** |
|
|
| Unlike many arithmetic datasets that focus only on direct |
| question-answer pairs, Arithmetic-XL follows a layered curriculum that |
| gradually introduces numerical concepts, multi-step computation, |
| operator precedence, compositional reasoning, adversarial examples, and |
| arithmetic embedded inside natural language. |
|
|
| Whether you're training a **5M parameter TinyLM** or a **multi-billion |
| parameter foundation model**, Arithmetic-XL provides a diverse |
| collection of arithmetic examples suitable for **pretraining**, |
| **continued pretraining**, and **specialized arithmetic fine-tuning**. |
|
|
| ------------------------------------------------------------------------ |
|
|
| # Why Another Arithmetic Dataset? |
|
|
| While training language models, I kept running into the same frustrating |
| problem. |
|
|
| The model could explain surprisingly complicated topics... |
|
|
| ...and then confidently tell me that: |
|
|
| > **17 × 8 = 121** |
|
|
| That felt slightly concerning. |
|
|
| I searched for arithmetic datasets that were: |
|
|
| - Large |
| - Diverse |
| - Procedurally verified |
| - Suitable for pretraining |
| - Rich enough to teach actual arithmetic patterns rather than simple |
| memorization |
| |
| I found some excellent datasets. |
|
|
| I also found datasets that contained the same question written five |
| hundred different ways. |
|
|
| After enough searching, I reached the conclusion every developer |
| eventually reaches: |
|
|
| > "Fine... I'll build it myself." |
|
|
| Several weeks, countless Python scripts, and far too many generated |
| equations later... |
|
|
| Arithmetic-XL was born. |
|
|
| ------------------------------------------------------------------------ |
|
|
| # What Makes Arithmetic-XL Different? |
|
|
| Arithmetic-XL is **not** simply a collection of equations. |
|
|
| It follows a structured curriculum inspired by how humans progressively |
| learn arithmetic. |
|
|
| Instead of immediately exposing the model to complicated expressions, |
| the dataset gradually increases difficulty through multiple specialized |
| layers. |
|
|
| The goal is to help models build numerical intuition before introducing |
| increasingly complex reasoning tasks. |
|
|
| ------------------------------------------------------------------------ |
|
|
| # Dataset Overview |
|
|
| ----------------------------------------------------------------------- |
| Property Value |
| ----------------------------------- ----------------------------------- |
| **Dataset Name** Arithmetic-XL |
|
|
| **Total Examples** **12,000,000+** |
|
|
| **Generation Method** Hybrid Procedural + LLM Assisted |
|
|
| **Verification** Computationally Verified |
|
|
| **Primary Language** English |
|
|
| **Domain** Arithmetic Reasoning |
|
|
| **Recommended Usage** Pretraining / Continued Pretraining |
| / Fine-tuning |
| |
| **License** MIT |
| ----------------------------------------------------------------------- |
|
|
|
|
