> ## Documentation Index
> Fetch the complete documentation index at: https://docs.llm-stats.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Creating Datasets

> Create benchmark datasets from the UI, git, or the Python SDK

## Recommended: git or UI

For most benchmarks, upload data via **git** or the **CSV importer** in the
benchmark hub. See [Uploading Data](/python-sdk/datasets/uploading-data) for
step-by-step instructions.

Git is best for:

* large or versioned datasets
* including scorer implementations and documentation alongside data
* reproducible, CI-friendly workflows

The browser editor and CSV importer are best for:

* quick prototypes
* small manually-curated test sets
* one-off imports

## SDK-based creation

You can also create and push datasets programmatically from Python. This is
useful for generating datasets from code or integrating with data pipelines.

### Create from Python data

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
import zeroeval as ze

ze.init()

dataset = ze.Dataset(
    "capital-cities",
    data=[
        {"row_id": "fr", "question": "Capital of France?", "answer": "Paris"},
        {"row_id": "de", "question": "Capital of Germany?", "answer": "Berlin"},
    ],
    description="Simple geography dataset",
)
```

<Tip>
  Add a stable `row_id` field whenever possible. It makes resume behavior and
  row-level tracking much more reliable.
</Tip>

### Create from CSV

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
dataset = ze.Dataset("/path/to/questions.csv")
print(dataset.name)  # e.g. "questions"
print(len(dataset))
```

The dataset name defaults to the CSV filename without the extension.

### Modify rows

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
dataset.add_rows(
    [{"row_id": "es", "question": "Capital of Spain?", "answer": "Madrid"}]
)

dataset.update_row(0, {"question": "Capital city of France?", "answer": "Paris"})

dataset.delete_row(1)
```

<Warning>
  Row updates by index can change row ordering and row content. Keep this in
  mind if your downstream logic assumes fixed order.
</Warning>

### Push to ZeroEval

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
dataset.push()
print(dataset.version_id, dataset.version_number)
```

After push, you can pin future runs to that version number:

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
latest = ze.Dataset.pull("capital-cities")
pinned = ze.Dataset.pull("capital-cities", version_number=dataset.version_number)
```

### Validate shape

```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
print(dataset.columns)  # sorted list of column names
print(len(dataset))     # number of rows
print(dataset[0])       # DotDict for first row
```