Tutorial

A guided, end-to-end tour of Firefly DataScience — from booting the app to serving a model.

This tutorial mirrors the runnable script samples/tutorial.py, which is covered by a test (tests/samples/test_tutorial.py), so everything here is guaranteed to work. It runs offline with no LLM key — the GenAI steps use deterministic stand-ins, and we show how to switch on a real LLM at the end.

uv add 'fireflyframework-datascience[tabular]'
uv run python samples/tutorial.py

We use a synthetic credit-risk dataset whose default risk is driven by debt-to-income — a ratio deliberately withheld from the model, so feature engineering has something real to discover.

The pattern every step rests on — the LLM proposes; the classical engine decides

Generative AI only ever proposes candidates here: feature code (step 4) and model choices (step 5). A deterministic classical engine cross-validates each one and a cost-benefit gate keeps it only if it measurably beats the current baseline. The LLM never touches the score — the data does. That is why the tour runs identically with or without an API key.

1. Boot the application

from fireflyframework_datascience import FireflyDataScienceApplication

app = FireflyDataScienceApplication.run()         # (1)!

1. Pass print_output=False to suppress the banner (the script does this so its test output stays clean).

This prints the banner and a wiring summary, loads configuration, builds the dependency-injection container, and discovers every adapter via entry-point auto-configuration. app.bean_count and app.applied_auto_configurations tell you what got wired.

Expected

On a fresh [tabular] install the container wires a couple dozen beans from roughly a dozen auto-configurations (exact counts depend on which extras are installed):

[1] App booted: 21 beans, 12 auto-configurations

See Architecture.

2. Build, load, and validate the data

The script generates the credit dataset with make_credit_dataset(). Default risk is a logistic function of debt_to_income = loan_amount / income, plus prior defaults and employment, but only the four raw columns are handed to the model — debt_to_income is the hidden driver feature engineering must rediscover.

import numpy as np
import pandas as pd

from fireflyframework_datascience.core.types import TaskType
from fireflyframework_datascience.datasets import Dataset
from fireflyframework_datascience.validation.adapters import BasicValidator


def make_credit_dataset(n: int = 800, seed: int = 11) -> Dataset:
    rng = np.random.RandomState(seed)
    income = rng.normal(60_000, 18_000, n).clip(15_000, None)
    loan_amount = rng.normal(18_000, 10_000, n).clip(1_000, None)
    employment_years = rng.uniform(0, 30, n).round(1)
    num_prior_defaults = rng.poisson(0.6, n)
    dti = loan_amount / income                                            # (1)!
    logit = -2.6 + 5.0 * dti + 1.3 * num_prior_defaults - 0.05 * employment_years + rng.normal(0, 0.25, n)
    default = (rng.uniform(0, 1, n) < 1.0 / (1.0 + np.exp(-logit))).astype(int)
    X = pd.DataFrame(
        {
            "income": income.round(2),
            "loan_amount": loan_amount.round(2),
            "employment_years": employment_years,
            "num_prior_defaults": num_prior_defaults,                      # (2)!
        }
    )
    return Dataset(
        "credit_applicants",
        X,
        pd.Series(default, name="default"),
        task=TaskType.BINARY,
        target_name="default",
        feature_names=list(X.columns),
    )


dataset = make_credit_dataset()
report = BasicValidator().validate(dataset.X, dataset.y)
assert report.ok                                  # no all-null columns, no null target, etc.
train, test = dataset.train_test_split(test_size=0.25, random_state=0)     # (3)!

1. dti drives the label but is never put into X — that is the signal step 4 has to recover.
2. Only these four raw columns reach the model; debt_to_income is deliberately absent.
3. train_test_split stratifies on the target for classification; here it yields 600 train / 200 test rows.

The BasicValidator catches empty data, all-null/constant columns, duplicate rows, and null targets before you waste time training.

Expected

[2] Data validated: ok=True

The dataset is 800 rows × 4 features, a TaskType.BINARY task; the split gives 600 train / 200 test.

See Datasets.

3. Classical AutoML

from fireflyframework_datascience.automl import AutoML

result = AutoML(cv=4).fit(train)
print(result.leaderboard_table())
print(result.evaluate(test))                      # holdout metrics

AutoML cross-validates each candidate trainer (RandomForestTrainer, LinearTrainer, HistGradientBoostingTrainer, plus the boosting libraries if installed), ranks them on a task-appropriate metric (roc_auc for binary), and refits the winner on the full training set. Each candidate is wrapped in an impute-and-scale preprocessing pipeline before scoring.

Expected

A leaderboard topped by linear, and a holdout roc_auc ≈ 0.85:

linear                   roc_auc=0.7867
random_forest            roc_auc=0.7493
hist_gradient_boosting   roc_auc=0.7335
EvaluationResult(primary=roc_auc=0.8498; accuracy=0.7900, f1=0.7853, precision=0.7851, recall=0.7900, roc_auc=0.8498, log_loss=0.4500)

Note

The leaderboard prints cross-validation scores on the training data, while evaluate(test) reports metrics on the untouched holdout — so the headline roc_auc (≈0.85) is higher than the CV figure (≈0.79). Both are real; they measure different things.

See Classical AutoML.

4. GenAI feature engineering

from fireflyframework_datascience.features import StaticFeatureProposer, FeatureProposal
from fireflyframework_datascience.features.genai import GenAIFeatureEngineer

proposer = StaticFeatureProposer([
    FeatureProposal("debt_to_income", "df['debt_to_income'] = df['loan_amount'] / (df['income'] + 1)", "DTI"),  # (1)!
    FeatureProposal("noise", "df['noise'] = 0.0", "should be rejected"),                                       # (2)!
])
engineered = GenAIFeatureEngineer(proposer, cv=4).engineer(train)
print(engineered.summary())

1. The hidden driver: its CV lift clears CostBenefitGate(min_gain=0.0), so it is accepted.
2. A constant column adds nothing, so the gate rejects it — the LLM never overrides that decision.

The loop is propose → execute (safely) → measure CV lift → gate. debt_to_income is accepted because it lifts the score; the constant noise feature is rejected. The LLM never decides — the measured score does.

Expected

GenAI feature engineering: 1 accepted, 1 rejected; roc_auc 0.7875 -> 0.7889 (lift +0.0013)

engineered.accepted lists debt_to_income; engineered.rejected lists noise with the reason no lift (0.7889 <= 0.7889). The lift is small but positive and real — the gate rejects anything that does not strictly beat the running baseline.

Static (no LLM)Agent (LLM)

StaticFeatureProposer stands in for the LLM so the tutorial runs offline with a fixed, reproducible set of proposals — exactly what the snippet above uses.

With a real model you swap in AgentFeatureProposer, which wraps a FireflyAgent and is built lazily (no LLM client is created at startup):

from fireflyframework_datascience.features.genai import AgentFeatureProposer, GenAIFeatureEngineer

proposer = AgentFeatureProposer(model="openai:gpt-4o")
engineered = GenAIFeatureEngineer(proposer, cv=4).engineer(train)

See Configuring the LLM.

See GenAI Feature Engineering.

5. The agentic ML-engineering loop

from fireflyframework_datascience.engineering import SequenceProposer, SolutionCandidate
from fireflyframework_datascience.engineering.loop import AgenticAutoML

proposer = SequenceProposer([SolutionCandidate("linear"), SolutionCandidate("random_forest"),
                             SolutionCandidate("hist_gradient_boosting")])
run = AgenticAutoML(proposer, cv=3, max_iterations=4).solve(train)        # (1)!
print(run.summary())

1. AgenticAutoML seeds the population, then reflects on the attempt history up to max_iterations times; a patience budget (default 3) stops the search once it stalls.

Each candidate is trained, cross-validated, and verified by a DeterministicVerifier — it must beat a trivial DummyClassifier(strategy="prior") baseline, not merely run (the "correctness ≠ ran" principle) — before the best one is selected. run.attempts is the full audited trail and run.valid_attempts are the ones that passed verification.

Expected

Agentic AutoML: 3 attempts (3 verified); best=linear roc_auc=0.7897 (baseline 0.5000)

All three seeded candidates clear the roc_auc=0.5000 trivial baseline, so all three are verified; linear wins.

See Agentic Loop.

6. Serve the model

from fireflyframework_datascience.serving import LocalModelServer

server = LocalModelServer()
server.load(result.best_model)
prediction = server.predict(test.X.iloc[[0]])     # score one applicant
print(int(prediction[0]))

LocalModelServer is the default, dependency-free server: it loads a fitted Model in the host process and answers predict / predict_proba. Heavier servers (e.g. BentoMLModelServer) live behind the serving extra.

Expected

The first holdout applicant is scored as a non-default:

[6] Served prediction for one applicant: default=0

See Serving & Lineage.

Turn on a real LLM

Steps 4 and 5 ran offline with deterministic stand-ins. To let a real model do the proposing, set your key and enable GenAI, then swap in the agent-backed proposers:

export OPENAI_API_KEY=sk-...                      # or ANTHROPIC_API_KEY=...
export FIREFLY_DATASCIENCE_GENAI__ENABLED=true
export FIREFLY_DATASCIENCE_GENAI__DEFAULT_MODEL=openai:gpt-4o   # or anthropic:claude-sonnet-4-5

from fireflyframework_datascience.features.genai import AgentFeatureProposer
from fireflyframework_datascience.engineering.loop import AgentSolutionProposer

Use AgentFeatureProposer in place of StaticFeatureProposer (step 4) and AgentSolutionProposer in place of SequenceProposer (step 5). Nothing else changes — the cost-benefit gate and the verifier still decide. The full guide, including providers, keys, cost gating, and secure execution, is in Configuring the LLM.

See also

• Quick Start
• Samples
• Configuration
• Agentic Loop
• Use Case: Lumen