Skip to main content
POST
Create dataset

Authorizations

Authorization
string
header
required

Create an API key in the console.

Headers

idempotency-key
string | null

Body

name
string
required
Required string length: 1 - 120
Pattern: ^[^\x00-\x1f\x7f]+$
examples
Example · object[] | null

Rows whose side we pick, by a hash of each context. Leave out training and evaluation.

training
Example · object[] | null

Rows to train on. A new dataset takes them only with evaluation rows.

evaluation
Example · object[] | null

Rows that only measure models. A new dataset of evaluation rows alone is for evaluation only.

evaluation_fraction
number | null

With examples: the share of contexts we hold out for evaluation, 0.1 to 1, and 0.2 when left out. 1 makes the dataset evaluation only.

Required range: 0.1 <= x <= 1
max_price_usd
string | null

The most you agree to pay for its first day of storage, which a dataset above the free size holds, in USD, such as the price you were shown. It is a decimal string, such as "2.500000000", never a JSON number, so no precision is lost: a number is refused with invalid_request, naming this field. If it costs more when it is admitted, it is refused with price_changed and nothing is held.

Pattern: ^\d{1,9}(\.\d{1,9})?$

Response

Successful Response

id
string
required
name
string
required
created_at
number
required
sha256
string
required

Names every row of the latest version: its first version's file digest, each later one's chained on.

version
integer
required

The latest version. Counts and the split are this version's.

rows
integer
required
contexts
integer
required
duplicates_removed
integer
required
method
enum<string>
required
Available options:
sft,
reward
archived_at
number | null
required

When we archived it: when credit couldn't cover its storage, or on a suspension. Restore it to start new work. Null in the library.

used_at
number | null
required

When work last started on it: a fine-tune, an evaluation or prompt tuning, or the labelling job that made it. Adding rows doesn't count. Null until then.

evaluation_only
boolean
required
evaluation_fraction
number
required

Its first version's evaluation share: as asked, or for a split of your own its evaluation contexts' share, which may be below the 0.1 a request takes.

split
Split · object
required
prompt_tuning
PromptTuningReadiness · object

What prompt tuning can tune in the training partition, counted when the dataset was made or grew, with the reasons the counts give worked out under the limits in force when it is read. Datasets of outcome scores have none.

input
InputReceipt · object
labeling
LabelingLineage · object
storage_window
StorageWindow · object | null

The day of storage held now, at the price it was accepted at, or null. Archiving or deleting it charges the minutes used of it; later days are priced as they start.

labels_to_check
integer

How many labels prompt tuning doubted in this dataset that nobody has dismissed, each row and decision once, as its labels-to-check route lists them.