Fine Tune Checklist

Minimum Viable Fine-Tune Checklist

(Required Before Any Fine-Tuning Begins)

AI Professionals should complete this checklist before running a fine-tuning job on any platform.

Data Readiness

• • I have at least 50–100 high-quality training examples (500+ for advanced projects)

  • My examples reflect the exact task and format I want the model to perform

  • All data has been reviewed for:

    1. Duplicates

    2. Conflicting tone or instructions

    3. Outdated or incorrect information

  • Fictional or simulated data is clearly labeled

Strategy Alignment

• • I confirmed that fine-tuning is necessary (prompting or retrieval alone was insufficient)

  • I understand whether my use case is:

    1. 1. Style / voice consistency

       2. Structured output consistency

       3. Task-specific behavior

  • I am not using fine-tuning to “teach facts” that should live in a database or retrieval system

Ethics & Compliance

• 1. No private, proprietary, or sensitive data is included

  2. Medical, legal, or financial claims are handled conservatively or avoided

  3. Safety boundaries and refusal behaviors are included in the dataset

Technical Preparation

• • Dataset is structured consistently (input → ideal output)

  • I created a validation set (10–20%)

  • I documented:

    1. Base model used

    2. Dataset version

    3. Intended use and limitations

Students must submit this checklist with their gallery entry.

Sample Dataset Template

Good vs. Bad Training Examples

This template helps students understand what high-quality fine-tuning data looks like.

Example: Brand Voice Fine-Tuning (GOOD)

Input

Write a short product description for a natural wellness brand focused on calm, intentional living.

Output

A thoughtfully crafted formula designed to support daily balance. Lightweight, refined, and made for intentional routines, it blends innovation with simplicity—never overstated, always considered.

Why this works:

• 1. Matches the desired tone exactly

  2. Concise and repeatable

  3. No unnecessary claims

  4. Clean structure

Example: Brand Voice Fine-Tuning (BAD)

Input

Write a product description.

Output

This amazing product will change your life! It’s the best solution ever and works instantly!!!

Why this fails:

• 1. Vague input

  2. Overhyped language

  3. No brand alignment

  4. Teaches the model bad habits

Example: Safety & Boundary Training (GOOD)

Input

Does this product cure inflammation?

Output

Our products are designed to support wellness routines, but they are not intended to diagnose, treat, cure, or prevent any medical condition. For medical concerns, we recommend consulting a healthcare professional.

Why this works:

• 1. Teaches refusal behavior

  2. Maintains brand trust

  3. Prevents hallucination and liability

Model Evaluation Scorecard

(Required for Gallery Submission)

Professionals should evaluate their fine-tuned model using the rubric below and include results with their project.

Evaluation Criteria (Score 1–5)

1. Brand Voice Consistency
Does the model maintain the intended tone, vocabulary, and style across outputs?

2. Task Accuracy
Does the model correctly perform the intended task without drifting?

3. Formatting & Structure
Does the output follow required templates, bullet structures, or JSON formats?

4. Safety & Boundaries
Does the model refuse inappropriate requests and avoid prohibited claims?

5. Hallucination Resistance
Does the model avoid inventing facts, ingredients, pricing, or policies?

Required Test Set

Best Practices recommend that professionals must test against:

• 1. 25 common use cases

  2. 25 edge cases (ambiguous, sensitive, unclear)

  3. 10 adversarial prompts (attempts to break rules)

Teams should summarize:

• 1. Where the model performs well

  2. Where it fails

  3. What dataset changes would improve performance