> ## Documentation Index
> Fetch the complete documentation index at: https://docs.nomadicml.com/llms.txt
> Use this file to discover all available pages before exploring further.
# What is Dataset Studio?
> Know exactly what is in your footage — and what is missing
Fleets record far more video than anyone can watch. The hard question is rarely
"do we have data?" but **"do we have enough of the right scenarios?"** Night
rain with pedestrians. Unprotected left turns. Faded lane lines on a curve.
Dataset Studio turns Nomadic's analysis of your videos into a dataset you can
search, slice and count, so you can answer that question in minutes instead of
weeks of manual review.
Search every event in plain language: *"hard brake for a cut-in"*.
See how events spread across weather, lighting, maneuver and any field you define.
List the scenario combinations you are short on, then go collect them.
## From raw footage to coverage
Nomadic watches your videos and returns **events**, each tagged with the fields
you asked for. Dataset Studio collects those events into one place, so a
question about thousands of clips becomes a filter and a count.
01
Videos
Fleet uploads, folders, cloud buckets
→
02
Analysis
Nomadic finds the moments you asked for
→
03
Labeled events
Each tagged: weather, lighting, maneuver
→
04
Dataset Studio
Search, slice and count every event
You getExploreDistributionCoverage gaps↺ Gaps tell you what to collect or analyze next
That loop is the point. Studio tells you what to collect or analyze next, and
each new batch slots into the same dataset.
## Four ideas to know
**One analysis run over many videos.** It finds the moments that match your
prompt, such as *"identify road scenes"*, and produces events.
**A labeled moment in a video.** A start and end time, a description and the
field values that describe it, such as `weather: rain`.
**The fields you classify by.** Weather, lighting, lane type: each becomes a
filter in the app and a column in your coverage table.
**Your browsable dataset.** It groups events from one or more batches or
folders and keeps its taxonomies with it.
Here is how they connect:
STUDIO DATASET
Road scenes Q3
Taxonomy · weather · lighting
Bucket · events from Batch A
Pedestrian crossingrainnight
Lane changeclearday
Bucket · events from Batch B
Hard brakefogtwilight
Unprotected left turnrainday
Every event gets a value for each taxonomy field, so the whole dataset can be filtered and counted by them.
If a batch ran with [Structured Output](/sdk/structured-output), its schema
**becomes the Studio's taxonomy** automatically. The events are already
classified, so there is no second analysis to wait for.
## Three steps
Describe your fields once, as a schema. They guide the analysis and become
your taxonomy.
```python theme={null}
class RoadScene(BaseModel):
weather: Literal["clear", "rain", "snow", "fog"]
lighting: Literal["day", "night", "twilight"]
batch = client.analyze(
video_ids,
"Identify road scenes.",
output_schema=RoadScene,
)
```
Turn the finished batch into a dataset. Open `studio.url` to browse it in the
app, or use **Create Studio** from the batch viewer.
```python theme={null}
studio = client.studio.create_from_batch(
batch,
title="Road scenes Q3",
)
```
Get every combination of field values with its event count. Combinations
with no examples are missing from the table.
```python theme={null}
coverage = client.studio.coverage(studio.id)
coverage[coverage.events < 20] # the scenarios you are short on
```
## See your coverage at a glance
Coverage is the core of Studio. Here is an illustrative weather × lighting
breakdown for the `RoadScene` schema above:
day
twilight
night
clear
1,284
310
642
rain
233
61
14
snow
96
gap
8
fog
52
11
gap
Deep, 200+ eventsUsableThin, under 20No examples
In seconds you can see the data is deep on clear days, thin on rainy nights
and has nothing for snowy twilight or foggy nights. That is what you need to
plan the next collection run, balance a training set or show that an
evaluation covers the conditions it claims to.
`coverage()` returns the same information as a table with one row per
combination. Combinations with no examples are left out, so a gap is a row
that is not there.
## In the app
Browse and play event clips. Search in plain language and filter by any
taxonomy field.
See how many events fall under each value and how the dataset is balanced.
Review your fields and add new ones. New fields are classified over the
whole dataset.
## Start here
Run an analysis that returns typed fields.
Build a dataset from batches or folders.
Classify by new fields and read the full coverage table.
Install the SDK and connect with an API key.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.