| name | fine-tune-readiness |
| description | Decide whether fine-tuning is justified versus prompting or RAG, and gate the training dataset before any LoRA/SFT/DPO run. Use when someone says 'let's fine-tune'. |
| version | 1.0.0 |
| author | Zorost Intelligence |
| license | MIT |
| metadata | {"zorost":{"tags":["fine-tuning","lora","sft","dpo","data-quality","build"],"related_skills":["eval-first-development","error-analysis-50","local-model-fit"],"program_weeks":[10]}} |
Fine-Tune Readiness
1 · Purpose
Fine-tuning is the last resort in a ladder of cheaper levers, this skill makes the
team climb the ladder in order, and gates the dataset before a single training run.
2 · When to use
- Whenever "we should fine-tune" is proposed.
- Before any SFT/LoRA/DPO run, to gate the dataset.
- When a fine-tuned model underperforms and nobody knows why (usually: the data).
3 · Inputs
- The task spec with metric and gate (
spec-first-ai-feature).
- An eval that can score the behavior change (
eval-first-development).
- The failure log from the current prompt/RAG system, clustered
(
error-analysis-50).
4 · Procedure
- Climb the ladder in order, recording the score each rung achieves:
(a) better prompting → (b) few-shot examples → (c) RAG with the right
documents → (d) fine-tuning. Quote the score at each rung; do not skip rungs.
- Name the failure class fine-tuning is meant to fix. Fine-tuning fixes
behavioral classes, format, tone, style, domain vocabulary, consistent
refusal/compliance patterns. It does not fix missing facts (that is RAG) or
weak reasoning (that is a bigger model or decomposition).
- If the failure class is knowledge or freshness, STOP: use RAG. If it is format
or tone and rungs a-c are exhausted, proceed.
- Gate the dataset before training:
- ≥ 200 examples for SFT format/behavior shifts (1,000+ for real moves);
preference pairs for DPO.
- Every example reviewed or generated against a written standard.
- No PII unless the compliance sign-off exists in writing.
- A held-out slice (≥ 10%) that training never sees.
- Choose the method: LoRA/QLoRA first (small adapter, reversible, cheap);
full fine-tune only when adapters demonstrably cannot move the metric. DPO only
after SFT, when you have preference pairs.
- Compute the hardware budget (
local-model-fit) or the cloud training cost
before launching. Write the number down.
- Train, then score on the held-out slice and the golden set. Compare against
the best prompt/RAG rung, not against the untrained base model alone.
- Ship only if the fine-tune beats the best non-trained rung by a margin that
justifies the serving and maintenance cost. Record the decision and the numbers.
5 · Anti-rationalization
| Excuse | Answer |
|---|
| "Fine-tuning is the professional way." | The professional way is the cheapest rung that clears the gate. Often that is a prompt. |
| "More data will fix it." | 200 reviewed examples beat 5,000 scraped ones. Data quality is the gate, not volume. |
| "It learned the format, so it works." | Format on training data is memorization. The held-out slice is the exam. |
| "Everyone fine-tunes now; it's cheap." | Training is cheap; owning a model's drift, evals, and retraining forever is not. |
6 · Red flags
- The proposal names a method (LoRA, DPO) before naming the failure class.
- The dataset was never read by a human.
- Success is defined against the base model, not against the best prompted system.
- No one can say what happens when the policy changes and the model must be retrained.
7 · Verify
- The ladder scores are recorded for prompting, few-shot, and RAG rungs.
- The dataset passed the four gates (size, standard, PII, held-out slice).
- The fine-tuned model's golden-set score and the best non-trained score are both
written down, and the ship/keep-prompting decision cites them.
8 · ZoroLogistics example
Week 10: the ops team wants replies in the company voice with strict section order.
Prompting gets 0.74 on the style rubric; few-shot 0.81; RAG adds nothing (the class
is behavioral, so the ladder says fine-tune). 600 SFT examples written against the
style standard, LoRA on a 14B at QLoRA, held-out slice 60. Fine-tuned: 0.93 vs.
few-shot 0.81, margin justifies the adapter. DPO follows with 300 preference pairs
for tone, lifting helpfulness without breaking format. Decision: ship the adapter,
keep the prompts under version control as the fallback.
© 2026 Zorost Intelligence LLC · zorost.com