Datasets

The Cohere platform allows you to upload and manage datasets that can be used for Fine-tuning models or in batch embedding with Embedding Jobs. Datasets can be managed in the Dashboard or programmatically using the Datasets API.

File Size Limits

There are certain limits to the files you can upload, specifically:

• A Dataset can be as large as 1.5GB
• Organizations have up to 10GB of storage across all their users

Retention

You should also be aware of how Cohere handles data retention. This is the most important context:

• Datasets get deleted 30 days after creation
• You can also manually delete a dataset in the Dashboard UI or using the Datasets API

Managing Datasets using the Python SDK

Getting Set up

First, let’s install the SDK

$
pip install cohere

Import dependencies and set up the Cohere client.

1
import cohere
2
3
co = cohere.Client(api_key="Your API key")

(All the rest of the examples on this page will be in Python, but you can find more detailed instructions for getting set up by checking out the Github repositories for Python, Typescript, and Go.)

Dataset Creation

Datasets are created by uploading files, specifying both a name for the dataset and the dataset type.

The file extension and file contents have to match the requirements for the selected dataset type. See the table below to learn more about the supported dataset types.

The dataset name is useful when browsing the datasets you’ve uploaded. In addition to its name, each dataset will also be assigned a unique id when it’s created.

Here is an example code snippet illustrating the process of creating a dataset, with both the name and the dataset type specified.

1
my_dataset = co.datasets.create(
2
    name="shakespeare",
3
    data=open("./shakespeare.jsonl", "rb"),
4
    type="chat-finetune-input",
5
)
6
7
print(my_dataset.id)

Dataset Validation

Whenever a dataset is created, the data is validated asynchronously against the rules for the specified dataset type . This validation is kicked off automatically on the backend, and must be completed before a dataset can be used with other endpoints.

Here’s a code snippet showing how to check the validation status of a dataset you’ve created.

1
ds = co.wait(my_dataset)
2
print(ds.dataset.validation_status)

To help you interpret the results, here’s a table specifying all the possible API error messages and what they mean:

Dataset Metadata Preservation

The Dataset API will preserve metadata if specified at time of upload. During the create dataset step, you can specify either keep_fields or optional_fields which are a list of strings which correspond to the field of the metadata you’d like to preserve. keep_fields is more restrictive, where if the field is missing from an entry, the dataset will fail validation whereas optional_fields will skip empty fields and validation will still pass.

Sample Dataset Input Format

{"wiki_id": 69407798, "url": "https://en.wikipedia.org/wiki?curid=69407798", "views": 5674.4492597435465, "langs": 38, "title": "Deaths in 2022", "text": "The following notable deaths occurred in 2022. Names are reported under the date of death, in alphabetical order. A typical entry reports information in the following sequence:", "paragraph_id": 0, "id": 0}
{"wiki_id": 3524766, "url": "https://en.wikipedia.org/wiki?curid=3524766", "views": 5409.5609619796405, "title": "YouTube", "text": "YouTube is a global online video sharing and social media platform headquartered in San Bruno, California. It was launched on February 14, 2005, by Steve Chen, Chad Hurley, and Jawed Karim. It is owned by Google, and is the second most visited website, after Google Search. YouTube has more than 2.5 billion monthly users who collectively watch more than one billion hours of videos each day. , videos were being uploaded at a rate of more than 500 hours of content per minute.", "paragraph_id": 0, "id": 1}

As seen in the above example, the following would be a valid create_dataset call since langs is in the first entry but not in the second entry. The fields wiki_id, url, views and title are present in both JSONs.

1
# Upload a dataset for embed jobs
2
ds = co.datasets.create(
3
    name="sample_file",
4
    # insert your file path here - you can upload it on the right - we accept .csv and jsonl files
5
    data=open("embed_jobs_sample_data.jsonl", "rb"),
6
    keep_fields=["wiki_id", "url", "views", "title"],
7
    optional_fields=["langs"],
8
    type="embed-input",
9
)
10
11
# wait for the dataset to finish validation
12
print(co.wait(ds))

Using Datasets for Fine-tuning

Once the dataset passes validation, it can be used to fine-tune a model. To do this properly, you must include at least five train examples per label.

In the example below, we will create a new dataset and upload an evaluation set using the optional eval_data parameter. We will then kick off a fine-tuning job using co.finetuning.create_finetuned_model.

1
# create a dataset
2
my_dataset = co.datasets.create(
3
    name="shakespeare",
4
    type="chat-finetune-input",
5
    data=open("./shakespeare.jsonl", "rb"),
6
    eval_data=open("./shakespeare-eval.jsonl", "rb"),
7
)
8
9
co.wait(my_dataset)
10
11
# start training a custom model using the dataset
12
co.finetuning.create_finetuned_model(
13
    request=FinetunedModel(
14
        name="shakespearean-model",
15
        settings=Settings(
16
            base_model=BaseModel(
17
                base_type="BASE_TYPE_CHAT",
18
            ),
19
            dataset_id=my_dataset.id,
20
        ),
21
    )
22
)

Dataset Types

When a dataset is created, the dataset_type field must be specified in order to indicate the type of tasks this dataset is meant for.

Datasets of type chat-finetune-input, for example, are expected to have a json with the field messages containing a list of messages:

1
{
2
    "messages": [
3
        {
4
            "role": "System",
5
            "content": "You are a large language model trained by Cohere.",
6
        },
7
        {
8
            "role": "User",
9
            "content": "Hi! What were Time magazines top 10 cover stories in the last 10 years?",
10
        },
11
        {
12
            "role": "Chatbot",
13
            "content": "Time magazines top 10 cover stories in the last 10 years were:\\n\\n1. Volodymyr Zelenskyy\\n2. Elon Musk\\n3. Martin Luther King Jr.\\n4. How Earth Survived\\n5. Her Lasting Impact\\n6. Nothing to See Here\\n7. Meltdown\\n8. Deal With It\\n9. The Top of America\\n10. Bitter Pill",
14
        },
15
    ]
16
}

The following table describes the types of datasets supported by the Dataset API:

Supported Dataset Types

Downloading a dataset

Datasets can be fetched using its unique id. Note that the dataset name and id are different from each other; names can be duplicated, while ids cannot.

Here is an example code snippet showing how to fetch a dataset by its unique id.

1
# fetch the dataset by ID
2
my_dataset_response = co.datasets.get(id="<DATASET_ID>")
3
my_dataset = my_dataset_response.dataset
4
5
# print each entry in the dataset
6
for record in my_dataset:
7
    print(record)
8
9
# save the dataset as jsonl
10
co.utils.save_dataset(
11
    dataset=my_dataset, filepath="./path/to/new/file.jsonl"
12
)
13
# or save the dataset as csv
14
co.utils.save_dataset(
15
    dataset=my_dataset, filepath="./path/to/new/file.csv"
16
)

Deleting a dataset

Datasets are automatically deleted after 30 days, but they can also be deleted manually. Here’s a code snippet showing how to do that:

1
co.datasets.delete(id="<DATASET_ID>")