Chapter 04: The Toy Task

Objective

Prove a text fine-tune pipeline on a tiny, unambiguous behavior: always answer as JSON with answer and confidence. Install mlx-lm, smoke-test it, then generate chat-format JSONL from a script.

Tutor pacing: concept → concept check → install → smoke test → write generator → emit data. Do not train in this chapter.

Concept: Why a toy task

Part 5’s flywheel will spin data → train → fuse → serve hundreds of times. Every defect multiplies. The most expensive bug in agent fine-tuning — loss masking — is invisible unless you deliberately test for it. So we use:

• a 1B-class model (minutes per loop, not hours)
• a behavior so trivial that success and failure are unambiguous

The behavior

Always respond with a JSON object containing keys answer and confidence.

Twenty-ish examples is enough to start. It is the “replace every number with BANANA” of fine-tuning: a habit the base model does not reliably have, checkable in one glance, zero domain knowledge.

Data format

mlx-lm expects a directory with train.jsonl, valid.jsonl, and optionally test.jsonl. Use chat format (what Part 3 needs):

{"messages": [{"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "{\"answer\": \"Paris\", \"confidence\": 0.99}"}]}

Generator, not a hand-written file

The script is the artifact; the JSONL is disposable. You will regenerate splits and sizes many times.

Concept check

Q1. Why prove the pipeline on “always return JSON with answer and confidence” instead of on a realistic coding agent task?

Q2. Why is a generator script the deliverable, not a carefully hand-edited train.jsonl?

Gate: Install mlx-lm

mkdir -p ~/pipeline-toy && cd ~/pipeline-toy
uv venv && source .venv/bin/activate
uv pip install mlx-lm

Gate: Verify it runs (smoke test)

Before writing datasets, prove generate works on the base model:

mlx_lm.generate \
  --model mlx-community/gemma-3-1b-it-4bit \
  --prompt "What is 2+2? Reply briefly." \
  --max-tokens 64

First run downloads weights. Pass: text prints without crash.

If this fails, stop — do not build JSONL yet.

Gate: Project layout (raw intent → train files)

mkdir -p ~/pipeline-toy/data
mkdir -p ~/pipeline-toy/scripts

You may keep a human-readable list of Q&A pairs in something like scripts/seed_questions.txt (raw), and have the generator write data/train.jsonl and data/valid.jsonl (train artifacts). Same idea as image raw/ → train/ in Part 1.

Gate: Write the generator and emit JSONL

Write scripts/make_json_dataset.py (or similar) that:

1. Defines ~20 user questions (or loads them from a seed file)
2. Emits assistant messages that are only JSON: {"answer": "...", "confidence": 0.0-1.0}
3. Writes data/train.jsonl and a smaller data/valid.jsonl

Run it. Show your tutor: wc -l data/*.jsonl and one sample line.

Do not hand-write the JSONL as the primary workflow.

Check your understanding

Q3. What must be true of the assistant side of each training example for this toy task?

Q4. You open the base model’s smoke-test reply and it is free-form prose, not JSON. Is that a problem before you train? Why or why not?

← Part 2 Index · Next: Chapter 05 →