LLM fine-tuning workflow template

Fine-tune a model step by step: define the task, build a clean dataset, train a LoRA adapter, evaluate against the baseline and ship.

Use this templateFree, opens in your browser. No sign-up.

LLM fine-tuning workflow template
Dataset checklist
Define the task + baseline eval (prompted model)
Collect 3,000 input → output examples
Clean: dedupe, strip PII, fix bad labels
Split 80 / 10 / 10: train, val, test
Train a LoRA adapter on the base model
Evaluate on the held-out test set
Beats baseline?
Add hard examples, fix label noise
Deploy behind a flag, A/B vs baseline
Starting hyperparameters
Setting
Start with
Why
Method
LoRA, rank 16, alpha 32
Trains < 1% of weights; cheap to iterate
Learning rate
2e-4, cosine decay
Lower if val loss jumps
Epochs
2–3
More epochs overfit small datasets
Batch size
16 (gradient accumulation)
Fits a single 24 GB GPU
Max sequence length
2,048 tokens
Covers 99% of examples
Examples look like real production inputs
Outputs written or approved by experts
Test set never seen in training
Version data + config with every run
LLM fine-tuning workflow
yes
no
iterate
The LLM fine-tuning workflow template as it opens in Kelvo: 9 boxes, 1 decision among them, 4 sticky notes, 1 table and 9 links.

What's on the board

Dataset checklist
Examples look like real production inputs · Outputs written or approved by experts · Test set never seen in training · Version data + config with every run

Table columns: Setting, Start with, Why.

Also on the board

  • Define the task + baseline eval (prompted model)
  • Collect 3,000 input → output examples
  • Clean: dedupe, strip PII, fix bad labels
  • Split 80 / 10 / 10: train, val, test
  • Beats baseline?
  • Train a LoRA adapter on the base model
  • Add hard examples, fix label noise
  • Evaluate on the held-out test set
  • Deploy behind a flag, A/B vs baseline
  • Starting hyperparameters

How to use this template

  1. Click Use this template. It opens in the Kelvo app as a new board, saved in your browser.
  2. Double-click any box to change its text. Replace the examples with your own content.
  3. Sketch, then convert. Draw boxes and arrows with the pen, select the strokes and click Convert sketch: they become real boxes and links.
  4. Ask Kelvo right on the board (Pro). With Pro, type @kelvo and what you want in any box or text, for example “@kelvo add error handling”, and press Enter. The change is made there and the request disappears; ⌘Z undoes it.
  5. Keep the prompt next to the diagram. Add a code block from the Insert (+) menu for prompts, schemas or config, right beside the boxes they belong to.

Good for

  • ai
  • llm
  • fine-tuning
  • lora
  • training
  • Explaining an LLM feature's architecture to engineers and stakeholders
  • Deciding between prompting, retrieval and fine-tuning
  • Planning evaluations and risks before launch

Related templates

  • Prompting vs RAG vs fine-tuning template previewPrompting vs RAG vs fine-tuning

    A decision tree for choosing prompt engineering, RAG, fine-tuning or an agent, with a cost and speed comparison.

  • RAG pipeline template previewRAG pipeline

    Retrieval-augmented generation end to end: ingest and embed documents offline, then retrieve, rerank and answer with citations.

  • LLM agent loop template previewLLM agent loop

    How a tool-using agent plans, calls tools, observes results and updates memory until the goal is met, with guardrails.

  • Prompt design canvas template previewPrompt design canvas

    Design a production prompt piece by piece — role, context, task, constraints, examples, output format — plus test cases.

  • Model evaluation plan template previewModel evaluation plan

    Metrics, datasets, targets and owners for evaluating an LLM or ML model, plus the gate a release must pass.

  • Skills matrix template previewSkills matrix

    Who knows what on the team, rated 1–4, so you can spot single points of failure and plan training.

All ai and machine learning diagram templates · All templates

Use the LLM fine-tuning workflow template