README.md

---
language:
- en
license: other
tags:
- synthetic
- math
- reasoning
- step-by-step
- text-generation
- language-modeling
task_categories:
- text-generation
task_ids:
- language-modeling
size_categories:
- 1M<n<10M
pretty_name: QuixiMath-1B
configs:
- config_name: preview
  data_files:
  - split: train
    path: preview/train-*.parquet
- config_name: 10M_tokens
  data_files:
  - split: train
    path: 10M_tokens/train-*.parquet
  - split: validation
    path: 10M_tokens/validation-*.parquet
- config_name: 100M_tokens
  data_files:
  - split: train
    path: 100M_tokens/train-*.parquet
  - split: validation
    path: 100M_tokens/validation-*.parquet
  - split: test
    path: 100M_tokens/test-*.parquet
- config_name: 1B_tokens
  data_files:
  - split: train
    path: 1B_tokens/train-*.parquet
  - split: validation
    path: 1B_tokens/validation-*.parquet
  - split: test
    path: 1B_tokens/test-*.parquet
train-eval-index:
- config: 10M_tokens
  task: text-generation
  task_id: language-modeling
  splits:
    train_split: train
    eval_split: validation
  col_mapping:
    text: text
- config: 100M_tokens
  task: text-generation
  task_id: language-modeling
  splits:
    train_split: train
    eval_split: validation
  col_mapping:
    text: text
- config: 1B_tokens
  task: text-generation
  task_id: language-modeling
  splits:
    train_split: train
    eval_split: validation
  col_mapping:
    text: text
---
# QuixiMath-1B

<img width="504" height="322" alt="math" src="https://github.com/user-attachments/assets/47fb7346-2ba1-49fb-b9e6-21917f078256" />

QuixiMath is brought to you by Eric Hartford and QuixiAI

https://github.com/QuixiAI/QuixiMath

## Dataset Summary

QuixiMath-1B is a synthetic math reasoning corpus generated from the QuixiMath
procedural problem generators. Each record contains a natural-language problem,
explicit step-by-step scratchpad opcodes, a canonical final answer, and metadata
for filtering or reweighting by skill, operation, grade band, and relative
difficulty.

The canonical corpus is coverage-first rather than prescriptively stratified:
trainers can choose their own sampling mix using the included metadata columns.
The size configs are nested prefix subsets within each split.

## How to Load

```python
from datasets import load_dataset

ds = load_dataset("QuixiAI/QuixiMath-1B", "100M_tokens")
train = load_dataset("QuixiAI/QuixiMath-1B", "100M_tokens", split="train")
```

## Configs And Splits

| Config | Split | Rows | Estimated tokens |
|---|---|---:|---:|
| `preview` | `train` | 50,000 | 6,134,016 |
| `10M_tokens` | `train` | 100,000 | 12,308,849 |
| `10M_tokens` | `validation` | 10,000 | 1,244,720 |
| `100M_tokens` | `train` | 800,000 | 98,641,588 |
| `100M_tokens` | `validation` | 50,000 | 6,164,238 |
| `100M_tokens` | `test` | 50,000 | 6,077,648 |
| `1B_tokens` | `train` | 8,800,000 | 1,104,706,100 |
| `1B_tokens` | `validation` | 100,000 | 12,333,425 |
| `1B_tokens` | `test` | 100,000 | 12,176,460 |

The largest config contains 9,000,000 rows and approximately
1,129,215,985 rough text tokens, estimated as `len(text) / 4`.

## Data Schema

Columns:

- `row_id`: stable integer row index within the split.
- `example_id`: stable string ID such as `train-000000123`.
- `problem_id`: generator-provided problem identifier.
- `generator`: generator class name.
- `generator_label`: generator class plus variant marker when applicable.
- `operation`: problem operation/category label.
- `grade_level`: one of `elementary`, `middle`, `high`, `college`, `graduate`.
- `difficulty`: integer 1-5, relative to `grade_level`.
- `problem`: problem text.
- `steps`: list of pipe-delimited scratchpad steps.
- `final_answer`: canonical answer string.
- `text`: training-ready text field containing problem, steps, and final answer.

## Dataset Stats

| Field | Value |
|---|---:|
| Default sampled skills | 509 |
| Default generator instances | 525 |
| Seed | 20,260,707 |
| Shard rows | 100,000 |

### Grade Distribution

| Grade level | Rows |
|---|---:|
| `college` | 2,963,901 |
| `high` | 2,324,185 |
| `graduate` | 1,539,789 |
| `middle` | 1,296,599 |
| `elementary` | 875,526 |

### Difficulty Distribution

| Difficulty | Rows |
|---|---:|
| `4` | 3,745,013 |
| `3` | 3,219,630 |
| `5` | 1,229,739 |
| `2` | 652,222 |
| `1` | 153,396 |

### Top Operations

| Operation | Rows |
|---|---:|
| `median` | 70,796 |
| `mean` | 70,303 |
| `multi_digit_subtraction` | 53,767 |
| `quantization_int8_affine` | 53,764 |
| `abacus_addition` | 53,757 |
| `kmeans_one_iteration` | 53,606 |
| `range` | 53,591 |
| `lu_decomposition` | 53,580 |
| `number_compare` | 53,568 |
| `multi_digit_addition` | 53,552 |
| `discrete_convolution` | 53,518 |
| `contour_integral_residue_theorem` | 53,514 |
| `mean_absolute_deviation` | 53,500 |
| `polynomial_add_sub` | 53,483 |
| `tensor_product_diagonal_apply` | 53,409 |
| `backprop_relu_step` | 53,383 |
| `systems_elimination` | 53,382 |
| `knn_classification` | 53,302 |
| `transportation_nw_stepping_stone` | 53,238 |
| `decimal_mul` | 53,199 |
| `dijkstra_trace` | 53,093 |
| `cramers_rule` | 52,974 |
| `classifier_precision_recall_f1` | 52,912 |
| `ratio_table` | 52,674 |
| `kernel_ridge_linear_2point` | 52,161 |

## Generation

Generated at: `2026-07-07T00:44:54.279574+00:00`

Source repository: `/home/hotaisle/datasets/QuixiMath`

Source git commit: `5283d55a85d7a127c8cfa5a1b5baf6b96dbc3301`

Source git dirty: `True`

Exact duplicate `(operation, problem)` pairs were skipped across the generated
largest splits before nested configs were materialized. Per-generator duplicate
and error counts are stored in `generation_stats.json`.

## Licensing Information

License: other