Algorithms From Scratch

How to Load Machine Learning Data From Scratch In Python

Suresh Madhra·Mar 10, 2026·7 min read

Before you model, you load

Every project starts the same way: a CSV file and a set of assumptions about what is inside it. Pandas hides that step behind one function call, which is wonderful right up until the moment a column silently becomes a string of objects and your model refuses to train.

In this tutorial you will load a CSV with nothing but Python's standard library, convert columns to the right types, and end up with a clean list of lists of floats — the format every other tutorial in this series expects.

1. Read a CSV file

The csv module gives you a reader that yields one list of strings per row. Two things matter: skip blank rows, and decide what to do with a header.

load_csv.py

from csv import reader
from io import StringIO

RAW = """sepal_length,sepal_width,petal_length,petal_width,species
5.1,3.5,1.4,0.2,Iris-setosa
4.9,3.0,1.4,0.2,Iris-setosa
7.0,3.2,4.7,1.4,Iris-versicolor
6.4,3.2,4.5,1.5,Iris-versicolor
6.3,3.3,6.0,2.5,Iris-virginica
5.8,2.7,5.1,1.9,Iris-virginica
"""

def load_csv(text, has_header=True):
    rows = [row for row in reader(StringIO(text)) if row]
    header = rows.pop(0) if has_header else None
    return header, rows

header, dataset = load_csv(RAW)
print("Header:", header)
print("Rows loaded:", len(dataset))
for row in dataset[:3]:
    print(row)

2. Convert strings to floats

Everything a CSV reader hands you is a string. Numeric columns need converting, and this is exactly where hidden whitespace and empty cells announce themselves.

str_to_float.py

def str_column_to_float(dataset, column):
    for row in dataset:
        value = row[column].strip()
        row[column] = float(value) if value != "" else None

dataset = [["5.1 ", "3.5"], ["4.9", ""], ["7.0", "3.2"]]
for i in range(2):
    str_column_to_float(dataset, i)
print(dataset)
print()
print("Note the None: missing values do not raise, they propagate. Decide now.")

3. Convert class labels to integers

Classification algorithms want integers, not strings. Build the lookup explicitly so you can map predictions back to human-readable labels later — an anonymous 0/1/2 in a report helps nobody.

str_to_int.py

def str_column_to_int(dataset, column):
    values = sorted(set(row[column] for row in dataset))
    lookup = {value: i for i, value in enumerate(values)}
    for row in dataset:
        row[column] = lookup[row[column]]
    return lookup

dataset = [["Iris-setosa"], ["Iris-virginica"], ["Iris-versicolor"], ["Iris-setosa"]]
lookup = str_column_to_int(dataset, 0)
print("Lookup:", lookup)
print("Encoded:", [row[0] for row in dataset])
print("Decode 2 ->", {v: k for k, v in lookup.items()}[2])

4. Load a real dataset over the network

Now put the three pieces together on the Pima Indians Diabetes dataset, fetched live in your browser.

load_diabetes.py

from csv import reader
from io import StringIO
from pyodide.http import pyfetch

URL = "https://raw.githubusercontent.com/jbrownlee/Datasets/master/pima-indians-diabetes.data.csv"

async def load_csv_from_url(url):
    resp = await pyfetch(url)
    text = await resp.string()
    return [row for row in reader(StringIO(text)) if row]

def str_column_to_float(dataset, column):
    for row in dataset:
        row[column] = float(row[column].strip())

dataset = await load_csv_from_url(URL)
print("Loaded %d rows and %d columns" % (len(dataset), len(dataset[0])))
for i in range(len(dataset[0])):
    str_column_to_float(dataset, i)
print("First row:", dataset[0])
print("Class distribution:",
      {c: sum(1 for r in dataset if r[-1] == c) for c in set(r[-1] for r in dataset)})

Checklist before you move on

Every column is the type you expect — print one row and read it.
Row count matches the source. A silent off-by-one usually means a header you forgot to skip.
Missing values have an explicit policy: drop, impute, or fail loudly.
The class column is last, encoded as integers, with a lookup you kept.

← Back to blog