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

# Data

> Upload your examples, check their format, and see which models they taught.

The **Data** screen, titled **Your examples**, holds the files your models learn
from. Every file gets a format check before you train on it. For tips on preparing
a good dataset, see the [Datasets](https://docs.valkyrieapp.azumo.com/guides/datasets.md) guide.

![Valkyrie Data screen listing uploaded example files](https://mintcdn.com/valkyrie-azumo/n5A0w3djfSi-SEkH/images/dashboard/data.png?fit=max\&auto=format\&n=n5A0w3djfSi-SEkH\&q=85\&s=e45b805c73c6afb1dfe0e6c517c4e794)

Your examples: every file with its format check and what it taught.

## The file list

Each row shows:

| Column | What it shows |
| - | - |
| **File** | The file name, its detected format, and its total number of rows |
| **Format check** | **Passed**, **N warnings**, **Needs fixing**, or **Not checked** |
| **Taught** | The model versions trained on this file, with their status |
| **Size** | The file size |

Use the icons at the end of a row to download or delete the file. **Upload data**
at the top right adds a new one.

## Start from an example set

No examples of your own yet? **Start from an example set** lists ready-made datasets
of about 2,000 examples each, enough for a fine-tune to learn the task and for its
test against the base model. Most come from public datasets that allow commercial
use; each card names its source and licence. Click **Copy to my data** and the set
becomes one of your own datasets.

![Valkyrie example datasets with Copy to my data buttons](https://mintcdn.com/valkyrie-azumo/_9aB329d4ICnAGkz/images/dashboard/data-starters.png?fit=max\&auto=format\&n=_9aB329d4ICnAGkz\&q=85\&s=e343673880670b6b1206319efb5d048f)

Example sets: copy one to see how a good dataset looks, or to try a fine-tune.

| Set | Format | What it teaches | Source |
| - | - | - | - |
| **Support replies** | Chat | Answer customers in your support team's voice, with the steps they need | [Bitext customer support](https://huggingface.co/datasets/bitext/Bitext-customer-support-llm-chatbot-training-dataset), CDLA-Sharing-1.0 |
| **Ticket triage** | Classification | Route incoming messages to the right team with one label (8 banking support queues) | [Banking77](https://huggingface.co/datasets/PolyAI/banking77), CC-BY-4.0 |
| **Invoice to JSON** | Instruction | Pull structured fields out of messy text, reliably as JSON | Generated by Valkyrie |
| **Plain English** | Instruction | Rewrite text in a clearer, simpler style | [ASSET](https://huggingface.co/datasets/facebook/asset), CC-BY-SA-4.0 |
| **General instructions** | Instruction | Follow everyday instructions well | [databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k), CC-BY-SA-3.0 |

## A dataset's page

Click a file to open it.

![Valkyrie dataset detail screen with format check and taught cards](https://mintcdn.com/valkyrie-azumo/n5A0w3djfSi-SEkH/images/dashboard/data-detail.png?fit=max\&auto=format\&n=n5A0w3djfSi-SEkH\&q=85\&s=8c830a7b8ff91863b2de4a4775a31926)

A dataset's page: format check, what it taught, and the next step.

* **Format check** shows the result, the **Detected format** (for example alpaca or
  chat), the **Columns** found, the **Rows checked** in the first part of the file,
  and the file's **Total rows**.
* **Taught** lists every model version trained on this file, with its status. Click
  one to open it in [Models](https://docs.valkyrieapp.azumo.com/dashboard/models.md).
* **Sample rows** shows the first examples in the file, as training will read them.
  Conversations show each message with its role; other formats show their columns.
* **Clean it up** and **Versions** are described below.

![Valkyrie dataset sample rows with text and label for each row](https://mintcdn.com/valkyrie-azumo/pEXaPnam1Ob9BH2B/images/dashboard/data-sample-rows.png?fit=max\&auto=format\&n=pEXaPnam1Ob9BH2B\&q=85\&s=44b66ad714d1d09f3963cb6606fb250c)

Sample rows: the first examples, as training reads them.

## Clean it up

Every cleanup saves a **new version** of the dataset and keeps the one you started
from, so nothing is ever overwritten.

![Valkyrie Clean it up card with cleanup actions](https://mintcdn.com/valkyrie-azumo/pEXaPnam1Ob9BH2B/images/dashboard/data-clean-up.png?fit=max\&auto=format\&n=pEXaPnam1Ob9BH2B\&q=85\&s=e81d727388be889d3959cf54dc1c7117)

Clean it up: each action saves a new version.

* **Remove duplicates** drops examples that repeat another one, ignoring differences
  in capital letters and spacing.
* **Remove examples that are too long** drops examples over 2,048 tokens, which
  training would cut off anyway.
* **Add paraphrased variants** writes a reworded copy of the question side of each
  example, keeping the answer as it is. It needs the
  [AI assistant](https://docs.valkyrieapp.azumo.com/dashboard/assistant.md), so it shows a padlock
  where the assistant isn't available.

When a dataset has more than one version, a **Versions** card lists the original and
every cleaned copy, with how many rows each has and what was removed.

The buttons at the top are **Download**, **Delete**, and **Fine-tune with these**,
which opens [New fine-tune](https://docs.valkyrieapp.azumo.com/dashboard/fine-tuning.md) with this file selected.

> [!NOTE]
>
> Deleting a file does not affect models you have already trained from it.

## Upload examples

Click **Upload data** to open **Upload examples**.

![Valkyrie Upload examples screen](https://mintcdn.com/valkyrie-azumo/n5A0w3djfSi-SEkH/images/dashboard/data-upload.png?fit=max\&auto=format\&n=n5A0w3djfSi-SEkH\&q=85\&s=2354d8fa84bcd0a8f49cc9291c2eef99)

Upload examples: drop a file and Valkyrie checks it right away.

**1. Choose a file**

Drop a file on the page or click **Choose file**. Valkyrie accepts JSONL, JSON,
CSV, Parquet, or text files up to 2 GB.

**2. Name it (optional)**

The name defaults to the file name.

**3. Upload**

Click **Upload**. Valkyrie checks the format as soon as the file lands and opens
the dataset's page with the result.

The page also shows **What a good example looks like**: one example per line, as a
conversation:

```json theme={null}
{"messages": [
  {"role": "user", "content": "..."},
  {"role": "assistant", "content": "..."}
]}
```

Instruction and response pairs, prompt and completion, plain text, and label columns
for classification are accepted too.

> [!TIP]
>
> No data yet? Download the ready-made sample, `customer-sentiment.jsonl` or
> `customer-sentiment.csv`, from the **No data yet?** card and upload it to try the
> whole flow.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.