Quick Start¶

Go from uv add to a fitted model and a working firefly-ds CLI in minutes — AutoML that fuses GenAI with classical ML and Deep Learning.

Firefly DataScience is a hexagonal, secure-by-default Python metaframework. The core stays import-light: heavy libraries (pandas, scikit-learn, XGBoost, MLflow, …) live behind optional extras and are loaded lazily, so you only install what you use. This page walks the shortest path: install an extra, boot the application, run AutoML two ways, and verify your environment with the CLI.

The reproducible pattern — the LLM proposes; the classical engine decides

Everything below is classical-first by default. GenAI is off unless you enable it, and when enabled it is a governed, cost-benefit-gated accelerator over a deterministic classical engine — never a black box. The defaults you boot with reflect that: genai disabled, sandbox = monty.

Install¶

Firefly DataScience requires Python 3.13+. The only hard dependency is the Firefly Agentic GenAI substrate; everything else is an optional extra. Pick the extra that matches what you are doing:

Core onlyClassical tabularAutoML stackGenAIEverything

No heavy ML libraries — just ports, the application bootstrap, and the DI container.

uv add fireflyframework-datascience

pandas, numpy, scikit-learn, xgboost, lightgbm, catboost, optuna.

uv add "fireflyframework-datascience[tabular]"

The curated bundle: tabular + tabfm + autogluon + tracking + validation + data.

uv add "fireflyframework-datascience[automl-stack]"

GenAI accelerators (script execution, embeddings, vector stores via Firefly Agentic).

uv add "fireflyframework-datascience[genai]"

tabular, DL, NLP, mlops, serving, lineage, orchestration, genai.

uv add "fireflyframework-datascience[full]"

Extras compose

Combine extras in a single brackets clause, e.g. uv add "fireflyframework-datascience[tabular,tracking,genai]".

Boot the application¶

FireflyDataScienceApplication mirrors the PyFly / Spring Boot lifecycle: load config → print banner → build the DI container → discover and apply auto-configurations → eagerly initialize singletons → return a ready ApplicationContext.

from fireflyframework_datascience import FireflyDataScienceApplication

# Construct and start in one call.
app = FireflyDataScienceApplication.run()  # (1)!

print(app.bean_count)                    # (2)!
print(app.config.default_ml_framework)   # (3)!
print(app.applied_auto_configurations)   # (4)!

run(**kwargs) constructs the application and immediately calls start(), returning a started ApplicationContext.
bean_count is the number of wired beans (len(app.container)).
The active ML framework from config — "sklearn" by default.
The auto-configuration classes whose conditions matched and were applied, in @order.

run(**kwargs) forwards to the constructor. Common options:

app = FireflyDataScienceApplication.run(
    config_dir="./config",        # directory containing firefly-datascience.yaml
    profiles=["local"],           # active configuration profiles
    print_output=False,           # silence the banner + wiring summary
)

When print_output is left on (the default), the application prints a short wiring summary after start — your profiles, beans, auto-config count, ml framework, genai state, and sandbox. Resolve wired components from the container by type:

from fireflyframework_datascience.models import TrainerPort

trainers = app.container.resolve_all(TrainerPort)   # all registered trainers
config = app.get(type(app.config))                  # or app.config directly

Run AutoML¶

AutoML is classical tabular AutoML: it validates the data (if a validator is wired), cross-validates a set of trainers (optionally tuning each), then fits the winner and ranks every candidate in a leaderboard. It works two ways — the framework serves both notebook-driven data scientists and DI-wired app developers.

Imperative (notebook style)¶

Build a Dataset from any sklearn-style dataset and call fit. With no arguments, AutoML() uses the default trainers (RandomForestTrainer, LinearTrainer, HistGradientBoostingTrainer), a default evaluator (SklearnMetricsEvaluator), and a default search policy (DefaultSearchPolicy).

import pandas as pd
from sklearn.datasets import load_breast_cancer

from fireflyframework_datascience.automl import AutoML
from fireflyframework_datascience.core.types import TaskType
from fireflyframework_datascience.datasets import Dataset

raw = load_breast_cancer(as_frame=True)
X: pd.DataFrame = raw.data
y = raw.target

dataset = Dataset(
    name="breast_cancer",
    X=X,
    y=y,
    task=TaskType.BINARY,                # breast cancer is a binary task -> roc_auc by default
    target_name="target",
    feature_names=list(X.columns),
)

# Cross-validate candidates and fit the winner.
result = AutoML(cv=5, n_trials=20, random_state=42).fit(dataset)

print(result.best_model.name)   # winning trainer
print(result.best_score)        # winner's CV score
for entry in result.leaderboard:
    print(entry)                # "<model>  <metric>=<score>"

# Predict with the fitted winner.
preds = result.predict(dataset.X)

The leaderboard is sorted best-first, and each entry stringifies as the model name padded to a column followed by <metric>=<score>:

Representative output

RandomForestTrainer
0.9789
RandomForestTrainer      roc_auc=0.9789
HistGradientBoostingTrainer roc_auc=0.9761
LinearTrainer            roc_auc=0.9743

Exact scores depend on the data, CV splits, and trial budget; the format is fixed.

fit accepts overrides — AutoML().fit(dataset, task=TaskType.REGRESSION, metric="r2") — otherwise the task comes from dataset.task and the metric from the evaluator's default for that task.

Hold out a test split the usual way:

train, test = dataset.train_test_split(test_size=0.25, random_state=42)
result = AutoML().fit(train)
report = result.evaluate(test)

Declarative (DI-wired)¶

AutoML.from_context pulls its trainers, evaluator, search policy, validator, and tracker straight from the application container, so an app's auto-configured (or custom) adapters are used automatically. Each component falls back to its default when not registered, and **overrides set the engine knobs (cv, n_trials, random_state).

from fireflyframework_datascience import FireflyDataScienceApplication
from fireflyframework_datascience.automl import AutoML

app = FireflyDataScienceApplication.run()

# Components are resolved from the DI container; kwargs override engine settings.
automl = AutoML.from_context(app, cv=5, n_trials=20)
result = automl.fit(dataset)

The `firefly-ds` CLI¶

Installing the package exposes the firefly-ds command (run with uv run firefly-ds <cmd>).

# Print the framework version.
firefly-ds version

# Check the environment and report which adapter extras are installed.
firefly-ds doctor

# Boot the app and list applied auto-configurations + registered beans.
firefly-ds introspect

# introspect with explicit config and profiles.
firefly-ds introspect --config-dir ./config --profile local --profile gpu

doctor verifies that the required Firefly Agentic substrate is present, then prints an installed / partial / not-installed status for every optional extra (tabular, tabfm, automl, dl, nlp, tracking, validation, featurestore, serving, lineage, orchestration, data, genai) — the fastest way to confirm your environment before a run.

Expected — firefly-ds doctor

Firefly DataScience doctor — v0.1.0
  python : 3.13.1 (macOS-15.5-arm64-arm-64bit)
  agentic: ok (required)
                   Optional adapter extras
┏━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━┓
┃ extra           ┃ status        ┃ modules ┃
┡━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━┩
│ tabular         │ installed     │ 7/7     │
│ tabfm           │ not installed │ 0/1     │
│ automl          │ not installed │ 0/1     │
│ dl              │ not installed │ 0/4     │
│ nlp             │ not installed │ 0/4     │
│ tracking        │ installed     │ 1/1     │
│ validation      │ installed     │ 1/1     │
│ featurestore    │ not installed │ 0/1     │
│ serving         │ not installed │ 0/1     │
│ lineage         │ not installed │ 0/1     │
│ orchestration   │ not installed │ 0/1     │
│ data            │ partial       │ 1/2     │
│ genai           │ installed     │ 2/2     │
└─────────────────┴───────────────┴─────────┘

installed means every representative module for the extra resolves; partial means some do; not installed means none. Your rows depend on which extras you installed.

If agentic is MISSING

The Firefly Agentic substrate is the one hard dependency. If doctor reports agentic: MISSING, the application will not boot — reinstall the package (the base install pulls Agentic in).