█
LastWrite
  • > Curriculum
  • > Pricing
  • > For Educators
  • > About
  • > Contact
Log InGet Started

Questions, concerns, bug reports, or suggestions? We read every message, write to us at [email protected].

More ways to reach us →
LastWrite

Structured computer science lessons for aspiring developers and security professionals.

[email protected]

(201) 785-7951

Mon–Fri, 9 AM–5 PM EST

Learn

  • Curriculum
  • Pricing

Company

  • About
  • For Educators & Schools
  • Contact Us

Legal

  • Terms of Service
  • Privacy Policy
© 2026 LastWrite. All rights reserved.
Curriculum/LLM Research and NLP/Fine-Tuning and Adaptation/Passion Project: Fine-Tune a Model
300 minAdvanced

Passion Project: Fine-Tune a Model

After this lesson, you will be able to: Fine-tune a small open model with LoRA/QLoRA on a task you curated, measure it honestly against the base model, and publish the adapter plus a reproducible recipe.

The Fine-Tuning capstone: a complete, reproducible adaptation experiment. This is the artifact that proves you can run real fine-tuning research on a student budget, and it feeds directly into the reproduction/paper in the Publishing sub-track.

Prerequisites:Compute Realities for Fine-Tuning

The build

Pick a narrow, measurable task (e.g. a specific classification or structured-extraction task, or a style/format the base model does inconsistently). Curate and decontaminate a dataset, fine-tune a small open model (1B-7B) with LoRA or QLoRA on a free/cheap GPU, and measure the fine-tuned model against the base model on a held-out test set with a metric appropriate to the task. The deliverable is the comparison, done correctly, not just a model that exists.

Milestones

Everything in a Git repo with a config-driven, seeded pipeline.

  1. 1

    Milestone 1: Define the task + the metric you'll report, and write the baseline (the base model, evaluated zero/few-shot) BEFORE training.

  2. 2

    Milestone 2: Curate the dataset; split train/val/test with a fixed seed; run + document decontamination.

  3. 3

    Milestone 3: LoRA/QLoRA fine-tune with tracked config, seed, and logged loss curves.

  4. 4

    Milestone 4: Evaluate base vs fine-tuned on the same test set with the same prompts/template; report the metric for both.

  5. 5

    Milestone 5: Run 3 seeds (or 3 data shuffles) and report mean + variance, not a single number.

  6. 6

    Milestone 6: Publish the adapter (Hugging Face Hub) + a README with the recipe, data description, decontamination note, and honest limitations.

ℹ️ The honesty bar

A fine-tune that beats the base model by a suspicious margin usually means leakage or an unfair baseline. Hold yourself to the standard the Publishing sub-track will demand: same eval for both models, decontaminated data, multiple seeds, and a limitations section. A modest, trustworthy improvement is worth far more than a large, fragile one, both for a paper and for what it teaches you.

How to talk about it

'I fine-tuned <model> with QLoRA on a task I curated, decontaminated the data against the test set, and measured a <X point> improvement over the base model across three seeds, with the recipe and adapter public.' That sentence demonstrates the full loop: data discipline, the method, honest evaluation, and reproducibility, which is exactly what a research advisor or applied-science interviewer is listening for.

Sign in and purchase access to unlock this lesson.

Sign in to purchase
←Compute Realities for Fine-Tuning
Back to Fine-Tuning and Adaptation
Fine-Tuning Job Readiness→