> ## Documentation Index > Fetch the complete documentation index at: https://docs.anyformat.ai/llms.txt > Use this file to discover all available pages before exploring further. # Building a dataset > Add documents, author ground truth, and slice with sub-datasets A **dataset** is the fixed set of documents an evaluation scores against. Each workflow has one, managed under **Health → Datasets**. This guide covers filling it, keeping its ground truth correct, and slicing it with tags. New to the terms? See [Datasets & evaluations](/concepts/evals) first. *** ## Adding documents There are two ways to put documents in a dataset — use whichever matches where your labelled data already lives. ### From your processed results Promote a document you've already run through the workflow: 1. Open its results and confirm every value is correct — **validate all of its datapoints**. The correct values become the file's [ground truth](/concepts/evals#ground-truth). 2. Choose **Add to dataset**. This **duplicates** the file into the dataset as an independent entity. From that point on, the dataset copy and the production file are separate — editing one never changes the other. A file can only be added this way once every datapoint is validated. Until then, **Add to dataset** stays disabled — you're establishing the correct answer, so it has to be complete. ### By direct upload Already have labelled data? Upload the documents and their expected answers straight into the dataset — no need to run them through the workflow first. 1. On the **Datasets** subtab, choose **Upload files**. 2. Drop your documents. To import expected values, include a **`.json` ground-truth file named after each document** — `invoice.pdf` pairs with `invoice.json`. 3. Optionally **tag** the uploaded files to slot them straight into a [sub-dataset](#sub-datasets-slicing-with-tags). 4. Review the staged list, then choose **Upload**. Ground truth is **optional** on upload — add documents alone and fill their values in later by editing the dataset table. Each `.json` is paired to its document by filename, and its values are checked against the workflow's current fields. The review step catches problems before anything is uploaded, flagging any file that is a **duplicate name** (dataset names must be unique — nothing is silently renamed), has **invalid JSON**, or is an **orphan** ground-truth file matching no document. A freshly added file — copied **or** uploaded — has ground truth but **no score yet**. It reads as ungraded until it's included in its first [evaluation](/guides/health/evaluations), which produces the prediction its ground truth is scored against. *** ## Editing ground truth The dataset's ground truth is the reference every evaluation is measured against, so keeping it correct matters more than anywhere else. Edit expected values directly in **Health → Datasets** — no need to go back to the original run. The **Ground truth** grid shows the expected values for each file, which are what an evaluation compares the workflow's output to. Editing ground truth changes **future** evaluations, not past ones. An evaluation you already ran keeps the score it recorded — run a new evaluation to measure the effect of your edits. *** ## Sub-datasets: slicing with tags One overall accuracy number hides where a workflow struggles. **Tag** files to create **sub-datasets** (slices) and read accuracy for each one separately. Tag by whatever distinction matters for your documents: * **Provider** — bank A vs bank B * **Document type** — invoices vs receipts * **Difficulty** — a "hard cases" tag for the documents that trip the workflow up You can then read results **overall** or **per sub-dataset** — for example *95% overall but 60% on hard cases* — and target the slice that needs work. Files with no tag make up the **Untagged** slice. When you [run an evaluation](/guides/health/evaluations), the **scope** you pick is either the full dataset or a single sub-dataset. *** ## Removing documents Select files in the **Datasets** grid and choose **Remove** to take them out of the dataset. Removal only affects the dataset — the production file is untouched. Removal is **final**: you can add the file to the dataset again, but that creates a **brand-new dataset entry** with a clean slate — it isn't linked to the removed one, and carries none of its earlier evaluation history or ground-truth edits. To measure the new entry, you'll need to run an evaluation over it. *** ## What's next? Score a workflow version against your dataset. How accuracy and confidence work across anyformat.