Build a tiny SFT dataset

The model is loaded; now it needs examples. We'll use a tiny, hand-built sentiment classification dataset so the whole pipeline is fast and inspectable. In a real project you'd have hundreds to thousands of rows; the mechanics are identical.

The shape: (prompt, completion)

From Track 1: an SFT example is a prompt paired with the target completion. For classification the completion is just the label. We keep the prompt instruction identical across examples so the model learns the task, not the wording.

raw = [
    {"prompt": "Classify the sentiment as positive or negative: I loved this movie.", "completion": "positive"},
    {"prompt": "Classify the sentiment as positive or negative: A complete waste of time.", "completion": "negative"},
    {"prompt": "Classify the sentiment as positive or negative: Best purchase I've made all year.", "completion": "positive"},
    {"prompt": "Classify the sentiment as positive or negative: It broke after one use.", "completion": "negative"},
    {"prompt": "Classify the sentiment as positive or negative: Absolutely delightful from start to finish.", "completion": "positive"},
    {"prompt": "Classify the sentiment as positive or negative: I want a refund.", "completion": "negative"},
    # ... in practice, hundreds more, covering the hard and edge cases
]

Wrap it in a datasets.Dataset

The datasets library gives us a fast, memory-mapped table with .map(), .filter(), and splitting — the standard container the Trainer expects.

from datasets import Dataset

ds = Dataset.from_list(raw)
print(ds)
print(ds[0])

Check class balance

Track 1 warned that imbalance lets a model cheat. Check it in one line:

from collections import Counter
print(Counter(ds["completion"]))   # e.g. Counter({'positive': 3, 'negative': 3})

Split into train and validation

We hold out a slice to watch for overfitting during training (Track 1). datasets makes this one call:

split = ds.train_test_split(test_size=0.2, seed=42)
train_ds, val_ds = split["train"], split["test"]
print(len(train_ds), "train /", len(val_ds), "val")

Keep a separate, hand-checked gold set too (Track 1, Lesson 8) — a handful of examples you will never train on, used in Lesson 2.7 to judge whether fine-tuning actually helped. With data in hand, the next step is turning these strings into the token tensors the model consumes.