Finetuning Method Selection

Fine-Tuning Method Selection

This is the router skill for the fine-tuning lifecycle: it decides whether fine-tuning is the right tool at all, and if so, which method and which base-model size class. Every other skill in this plugin assumes this routing already happened — start here before opening lora-qlora-recipes, preference-optimization, or grpo-rlvr-training.

When to Use This Skill

• Starting any fine-tuning effort, before a framework or base model has been chosen.
• Unsure whether RAG or prompt engineering would solve the problem more cheaply than training.
• Choosing between preference optimization (DPO family) and a reinforcement method (GRPO/RLVR) for the same underlying task.
• Sizing a candidate model/method combination before committing to a run.

Quick Reference

Off-Ramps First

Most requests that sound like “fine-tune this” are served better and cheaper elsewhere. Check these off-ramps before opening a training run:

• Knowledge-bound and volatile (the gap is facts that change — prices, docs, current events): route to RAG, not fine-tuning. A fine-tuned model bakes in a snapshot; volatile facts go stale immediately.
• Behavior-bound and shifting (the desired behavior is still being figured out, or changes per request): route to prompt engineering. Fine-tuning locks in a behavior; don’t lock in one that hasn’t stabilized yet.
• Stable, dense domain knowledge: this is where continued pretraining (CPT) enters, sized by how much domain text exists:

CPT learning rate ≈ 10% of the pretraining LR. CPT is guidance-only in this plugin — sizing and LR guidance live here, but this plugin does not execute a CPT run.

Method Router

Once the off-ramps are ruled out, this is the full decision tree (verbatim from the research this plugin is built on):

New FACTS?  volatile → RAG | stable+dense → CPT (LR ~10% of pretrain) → SFT
New BEHAVIOR? shifting → prompt-engineering | stable:
  demos → SFT (LoRA/QLoRA, all-linear, α=2r)
  preference pairs → DPO (SimPO if length-bias, ORPO if memory-bound)
  unpaired 👍/👎 → KTO
  verifiable success → RLVR + GRPO (DAPO/GSPO/Dr.GRPO per failure mode)
Deploy: FP8 (Hopper+) | NVFP4 (Blackwell scale) | AWQ (older) | GGUF+imatrix (edge)
BEFORE ANY OF THIS: the eval harness must exist first.

Read the tree top-down: answer “new facts or new behavior,” then follow the branch that matches the data shape in hand (demos, preference pairs, thumbs up/down, or verifiable success/failure). The data shape picks the method — not the other way around.

Worked Routing Examples

• “Users want the assistant to follow our support macros exactly.” Behavior is stable and demonstrable from transcripts → demos → SFT.
• “We have pairs of good/bad responses from reviewer thumbs-up/down, unpaired.” → unpaired signal → KTO, not DPO (DPO needs paired preferences).
• “The model can already solve some of these math problems and we can grade correctness automatically.” → verifiable success signal → GRPO+RLVR, and only after confirming the model succeeds at least sometimes (see Key Routing Facts below).
• “We want the model to know this week’s pricing page.” → volatile facts → RAG, no training run at all.

Key Routing Facts

• Loss-function choice is low-leverage. A 240-H100-run study found method choice worth ~1 percentage point versus ~50 points for model scale, and zero of 20 DPO variants beat vanilla DPO. Don’t spend a routing decision agonizing over DPO-variant selection — spend it on getting the data shape and scale right.
• DPO is for taste, GRPO+RLVR is for reasoning. Preference pairs that encode a subjective judgment (tone, style, “which answer is better”) route to DPO. Tasks with a verifiable pass/fail signal (math, code, tool calls) route to GRPO+RLVR instead.
• RL is not the fix for a model that never succeeds. GRPO and other RL methods sharpen an existing capability — they don’t teach one from zero. If the model doesn’t yet understand the task or output format, run SFT first; only bring in RL once the model succeeds at least sometimes.

Common Routing Mistakes

• Reaching for fine-tuning to fix facts that change weekly — that’s a RAG problem, and fine-tuning will just go stale faster than the source data does.
• Picking a DPO variant before checking whether the actual bottleneck is data quality or model scale — variant choice is the ~1pp lever, not the ~50pp one.
• Starting an RL run on a model that fails every rollout — route to SFT first so RL has something to sharpen.
• Treating CPT as the default for “the model doesn’t know our domain” — check the data volume thresholds first; under 500MB, RAG or RAG+fine-tune iterates faster than a CPT run.

Model Selection

Base-model choice is size-class first, family second, and it goes stale fast — so it lives in exactly one place: references/model-catalog.md. That file is the only place in this plugin (and in the DGX Spark ops plugin) that names a base model family. Neither this skill nor references/memory-math.md names one; both describe models by size class only (for example, “8B-class LoRA,” not a model name).

The catalog is dated on purpose — model rankings turn over quarterly. It carries a “last verified” date and a refresh checklist. Before trusting a row, check that date; if stale, work the refresh checklist in the catalog before recommending a model from it.

Precedence when the catalog and a method skill disagree: the catalog’s per-row Notes column states hardware/size-class feasibility, not a method recommendation — lora-qlora-recipes‘s LoRA vs QLoRA vs Full FT table (routed by task shape) governs the actual method choice.

Memory Feasibility

Before committing to a method, size it: total memory ≈ params × dtype bytes + optimizer state + gradients + activations. Work each term for the chosen dtype and method (full fine-tune, LoRA, or QLoRA) — worked worksheets and size-class examples live in references/memory-math.md.

On DGX Spark specifically, unified-memory behavior breaks the naive estimate (transient load peaks, nvidia-smi underreporting, thermal throttling on long runs). Once the dgx-spark-ops plugin is installed, defer Spark-specific feasibility calls to its spark-memory-thermal-ops skill rather than re-deriving them here.

Related Skills

Once this skill has picked a method, hand off to the skill that executes it:

• lora-qlora-recipes — SFT via LoRA/QLoRA
• preference-optimization — DPO, ORPO, KTO
• grpo-rlvr-training — GRPO with verifiable rewards

No method is selected before the eval harness exists — see eval-harness-first.