Argilla hosting

Human feedback and dataset curation for LLM work.

ML platformApache-2.0from $11.99/mo

In short

Argilla hosting at AIHostBox means a dedicated, single-tenant Argilla instance on our infrastructure: we install it, pin it to a reviewed release, put it behind TLS on your domain, back it up nightly and keep it patched. Plans start at $11.99 per month and the admin account, the data and any API keys are yours.

Feedback collectionDataset curationAnnotator UIHF integration

Argilla is an open-source platform for building datasets with humans in the loop, written by the Argilla team that joined Hugging Face. You define a dataset with fields and questions (labels, multi-label, ratings, rankings, free text and spans) through the Python SDK, then annotators work through the records in a web UI. Model suggestions, metadata filters and vector search help them move faster and focus on the records that matter.

Finished datasets can be exported to the Hugging Face Hub or loaded straight into training code. Argilla is not a single process: it is a FastAPI server, a background worker, PostgreSQL and Elasticsearch, which is more than most laptops should run all day. A hosted instance gives annotators a stable address and keeps their work in one backed-up place. Several annotators can label the same dataset, and their responses are stored side by side for agreement checks.

Argilla hosting plans

PlusRecommended minimum

4 GB RAM · 2 vCPU · NVMe

$11.99/mo

Order

ProTeams & heavier load

8 GB RAM · 3 vCPU · NVMe

$18.99/mo

Order

Every plan is a dedicated single-tenant instance with NVMe storage, nightly off-box backups, TLS and managed patching. Quarterly, semi-annual and annual terms are discounted at checkout. Model tokens are not included — you bring your own key.

How we run Argilla

The official compose stack: the argilla/argilla-server container plus its background worker, PostgreSQL, Elasticsearch 8 and Redis, with separate volumes for Argilla, Postgres and Elasticsearch.

What we configure for you

USERNAME, PASSWORD and API_KEY replaced from the compose defaults before first start
ARGILLA_DATABASE_URL, ARGILLA_ELASTICSEARCH and ARGILLA_REDIS_URL on a private network
ES_JAVA_OPTS heap sized to the plan
BACKGROUND_NUM_WORKERS matched to the vCPUs
Hugging Face OAuth sign-in on request, with your OAuth app credentials

Which plan fits

Plus (4 GB) runs the full stack with Elasticsearch on a small heap, enough for a few annotators and datasets of tens of thousands of records. Pro (8 GB) gives Elasticsearch room for larger datasets, vector fields and more concurrent annotators.

The honest limits

Elasticsearch is the most memory-hungry part and the usual cause of trouble on small machines, so 2 GB is not offered. No new features are planned upstream.

What people run Argilla for

Fine-tuning data

Turn corrections into a training set.

Quality review

Sample production output for human scoring.

Preference data for RLHF and DPO

Rank or rate paired model responses to build preference datasets.

Licence and version policy

Item	Position
Upstream licence	Apache-2.0
Sold as managed hosting	Yes
Tenancy	Single tenant — one customer per instance
Patching	Critical advisories within 24 hours, high severity within 72

Our full position on licences, including the applications we decline to sell, is on the licensing page.

Works well with Argilla

More ML platform we host

Questions

Argilla hosting: your questions

Is Argilla still developed?

Upstream describes the codebase as mature and stable: the original authors have moved on, and no new features are planned, but bug fixes and patches continue. That makes it a predictable tool to build a labelling process on, though you should not expect new capabilities.

How do annotators log in?

The owner account is created at deployment, and further users are added with owner, admin or annotator roles and assigned to workspaces. Argilla also supports signing in with Hugging Face accounts through OAuth, which requires registering an OAuth app on Hugging Face.

How do I get data in and out?

Through the argilla Python SDK: create a dataset with rg.Settings, log records, then export responses to a pandas-friendly format, to disk or to the Hugging Face Hub. Datasets can also be imported from the Hub directly, which is the easiest way to migrate from another Argilla server.

Which Argilla version do you run?

Argilla 2.x, the current line with the rg.Dataset API. Datasets built with the 1.x FeedbackDataset already use the 2.x format, but the older text classification, token classification and Text2Text datasets need migrating with the 1.x compatibility module in the new SDK.

What does the backup cover?

Users, workspaces, datasets and responses are stored in PostgreSQL, while Elasticsearch holds the search index built from them. Both volumes are included in the nightly off-box backup, and the PostgreSQL data is the part that matters most for a restore.

Run Argilla without running a server

A dedicated Argilla instance, deployed and maintained by us. Month to month, cancel whenever.

Browse the catalogue Talk to a human