Welt Python SDK

From a DataFrame to a prediction. Use familiar Classifier and Regressor estimators backed by durable predictors in your private Welt workspace. Install welt-client, import welt. Python 3.11–3.13.

Start here · Guides · Reference · Models and limits

Install

python -m pip install "welt-client[parquet]"

Run welt login and approve your workspace in the browser. This saves a private, origin-bound credential for later Python sessions. In a notebook, connect once:

from welt import Client

client = Client().connect()

Open its printed verification link on your own computer if Python is remote. Python connection stays in this process; use save=True explicitly to persist it. Constructors never open a browser. Explicit keys and WELT_API_KEY still work and have precedence over connected/saved credentials. Connection guide.

Modelling and research downloads may consume workspace credits. Check Workspace → Billing for available credits, pending reservations and spend caps before running the examples. Usage and recovery explains accounting; payments and account changes require browser authorization.

Learn with real data

Download Iris from our research inventory, look at the table, split train/test, fit a classifier, predict unseen rows, and measure the result:

from pathlib import Path
from tempfile import mkdtemp
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, classification_report
from welt import Client, Classifier

# 1. Get a real table from Welt's research inventory.
dataset_id = "asset-d839644b6de36fc5cafdda18c5daeca3"
path = Path(mkdtemp(prefix="welt-iris-")) / "iris.parquet"
client = Client()
info = client.research_dataset(dataset_id)
print("Dataset:", info["name"])
client.download_research_dataset(dataset_id, path)
data = pd.read_parquet(path)
target = "target"  # The outcome column in this table.

# 2. Look at the rows before modelling.
print(data.head())
print("Rows and columns:", data.shape)
data.info()
print(data[target].value_counts())

# 3. Set aside test rows before fitting; never train on their answers.
train, test = train_test_split(
    data, test_size=0.2, random_state=9, stratify=data[target])
query = test.drop(columns=target)
y_test = test[target]
print("Training rows:", len(train), "— Test rows:", len(test))

# 4. Create a model and train it on the training table.
model = Classifier(model="tabicl-v2", random_state=9)
model.fit(train, target=target)

# 5. Predict the held-out rows, without passing their target column.
predictions = model.predict(query)
print(pd.DataFrame({"actual": y_test.to_numpy(), "predicted": predictions}).head())

# 6. Compare predictions with the answers we kept aside.
print("Test accuracy:", accuracy_score(y_test, predictions))
print(classification_report(y_test, predictions, zero_division=0))
print("Keep this predictor ID:", model.predictor_id_)

Accuracy and per-class precision/recall/F1 describe your held-out run, not a promised benchmark score. Start explains each step and teaches numeric predictions with real Yacht data. fit uploads only your training rows; your workspace retains the resulting dataset and predictor for reuse.

Research metadata and one-step loading

Keep the catalogue detail (schema, column descriptions and provenance) beside a download with metadata=True, or load a table directly with the research extra (pip install "welt-client[research]"):

saved = client.download_research_dataset(
    dataset_id, path.with_name("iris-2.parquet"), metadata=True)
# saved.path, saved.metadata_path ("iris-2.parquet.metadata.json"), saved.metadata
data, meta = client.load_research_dataset(dataset_id, metadata=True)
print(meta.name, meta.column_descriptions)

Without metadata=True the download returns its local path. Omitting version selects the currently published file; the SDK verifies its size and checksum. load_research_dataset removes its temporary file after loading. See Research datasets.

The released client includes sync/async resources, resumable file uploads, durable batch jobs, workspace selection and native causal objects. Estimator CSV/Parquet convenience reads load files into memory; Client.upload_file uses upload sessions and supports recovery with a retained upload ID. Available formats, models, tasks and limits depend on the connected server. Inspect its catalogue before executing. Each TabICLv2, Kumo Tabular Medium, TabDPT 1.3 and Mitra-v2 classification and regression profile has measured shapes: TabICLv2 and Kumo Tabular Medium up to 8,000 training rows × 20 features (or 4,000 × 40, 2,000 × 80), Mitra-v2 up to 4,000 × 20, TabDPT 1.3 up to 500 × 20. Direct predictions take up to 100 rows per call, and classification up to 10 classes. Client().models() reports the current limits. Capabilities explains scope.

A local wait timeout leaves server work running. Reconnect by its job ID rather than fitting again; recovery describes typed failures, authentication, cancellation and retry behavior. Probabilities are not calibrated coverage. ArrowFM, CDFM and AVICI discovery preserve native graph and score semantics; causal conversion is restricted to declared native DAGs and needs the welt-client[ergodic] extra; ergodic then identifies and estimates effects locally. The complete all-model catalogue remains gated.

Examples and development

Runnable script · Jupyter notebooks · Research datasets · Contributing

uv sync --frozen
uv run --frozen pytest -q
uv build

Offline tests patch HTTP transport outside user examples; they neither contact the hosted service nor measure foundation-model quality. Notebook sources have cleared outputs; clean-wheel execution and bounded authorized live acceptance are separate evidence. CI requires no private backend, credentials or checkpoints.

Agent workflows, workspace selection and the released command reference: CLI.

Explicit hidden-model runs

client.model("chronos-2") inspects an explicit hidden ID while client.models() keeps the main catalogue. client.submit_model_run(payload, idempotency_key="your-key") returns a durable Job; job.result() returns the native task result. AsyncClient has matching methods. The payload follows the API's tabular support/query, causal table or univariate forecasting contract; check the model metadata for enabled tasks, limits, readiness and workspace access.

from welt import Client
client = Client()
job = client.submit_model_run({
    "model_id": "chronos-2", "task": "univariate_forecasting",
    "series": {"values": [float(i) for i in range(48)],
               "start": "2026-01-01T00:00:00Z", "frequency_seconds": 3600},
    "horizon": 12,
}, idempotency_key="forecast-example-1")
result = job.result()
print(result["predictions"])

CLI equivalents: welt models get chronos-2, welt run --input request.json, welt jobs wait JOB_ID. Use welt model-runs delete JOB_ID --yes or client.delete_model_run(id) to explicitly remove a terminal run's input/result. Deleting keeps the idempotency/usage record; a new run needs a fresh key. Native model libraries remain server-side. Restricted models need an internal testing workspace; hidden status alone does not grant execution access. TabSTAR's trained state stays alive for the single run, and these runs do not publish reusable support-only fitted predictors. LimiX query splitting can change its transductive results. API errors, cancellation and timeouts retain existing typed semantics.

Compare models with cross-validation

client.evaluate(data, "target", task="classification") cross-validates every eligible model (or up to 8 you name with models=[...]) on the same seeded folds and returns an Evaluation: summary() is the leaderboard against a baseline, folds() and predictions() are DataFrames, show() displays the HTML report in Jupyter and report_link() creates a signed link anyone holding it can open for up to seven days. evaluation.fit() then fits the chosen model on the full dataset. It requires SDK 0.12 or later and a server release with evaluations; an older server raises NotFoundError with code="evaluations_unavailable". Evaluations guide.

Forecast time series

Forecast(model="chronos-2") forecasts one series or a long-format panel with Chronos-2. Fit with time, an optional id column, one or more target columns, and past_covariates, future_covariates and categorical drivers; then call forecast(horizon, future=known_future_table). One request covers every series and target, and to_frame() returns one row per series, target and step with the median and q_<level> quantiles. Limits come from the model's published envelopes (client.model("chronos-2")["forecast_envelopes"]), and from the forecasting release a forecast is priced at the prediction rate per token: Σ (history rows + horizon) × (targets + covariates), rounded up. Quantiles are not calibrated intervals; crossed native quantiles are sorted and counted in rearranged_steps. Panels and covariates need SDK 0.14 and a server with the forecasting task. History and future can also be workspace dataset IDs; tables are sent inline, never uploaded automatically, and bodies above 10 MiB are refused before sending. client.delete_forecast(id) removes a finished forecast's input and result. The univariate predict API from SDK 0.11 stays supported. Forecast(model="timestral-209m") (or timestral-28m), listed after Chronos-2, uses the same calls and price but forecasts one series per request: SDK 0.14.1 refuses a panel before sending with unsupported_capability, so fit and forecast each series separately. Timestral weights: 2026 - Mistral AI, Modified MIT licence (Timestral-209M, Timestral-28M) (forecast with Timestral). Forecast guide · tutorial notebook · Use Chronos-2 inside ergodic's backtest and reconcile: https://docs.ergodic.dev/forecast-with-welt.html