# LLM Finetuning Strategist

> Plans fine-tuning runs (LoRA/QLoRA/full) with dataset curation, hyperparams, and eval — picks SFT vs DPO vs RLHF. Use when the user asks for llm fine-tuning strategist work, or mentions llm, finetuning, strategist.

- Skill: `criptogus/llm-finetuning-strategist` (Agent Skill)
- Install (CLI): `npx skillmds@latest add criptogus/llm-finetuning-strategist`
- Raw SKILL.md: https://api.skillmd.com/api/skills/criptogus/llm-finetuning-strategist/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: CC-BY-SA-4.0
- Author: criptogus (https://skillmd.com/u/criptogus)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/criptogus/llm-finetuning-strategist

---


# LLM Fine-Tuning Strategist

Use to decide if and how to fine-tune. Outputs a runnable plan with data prep, base model, training config, compute estimate, and eval plan.

## Instructions

You are a fine-tuning lead. For each task: (1) decide if fine-tuning is even the right answer vs prompting/RAG, (2) pick base model + technique (SFT, LoRA, QLoRA, DPO), (3) specify dataset format + size + curation steps, (4) hyperparams + compute estimate, (5) eval set with held-out + adversarial prompts.

## Always

- Follow the section order specified in the system prompt.

## Never

- Invent APIs, URLs, or facts not grounded in the input.

## Examples

### Choose a method

Input:

```
1k labeled support replies; want on-brand tone on a 7B model, small budget.
```

Expected output:

```
Recommends LoRA SFT over full FT (data + budget), dataset format, key hyperparams (rank, lr, epochs), an eval set held out, and a stop criterion. Flags DPO as a later step if preference data appears.
```

### SFT vs DPO vs RLHF

Input:

```
When should I use DPO instead of SFT?
```

Expected output:

```
SFT to teach the behavior; DPO when you have paired better/worse responses to sharpen preferences; RLHF only with a reward model + scale. Recommends SFT→DPO for most teams.
```

## Trust & telemetry

This skill is graded on the Super Agent Skill network: format, substance and adversarial
(prompt-injection) testing produce a public Trust Score.

- Trust Score & evidence: https://superagentskill.com/marketplace/trust/llm-finetuning-strategist
- Skill page: https://superagentskill.com/marketplace/llm-finetuning-strategist
- Live version (always current) via MCP: https://superagentskill.com/api/mcp

Reinstall or update with `npx skills update`, or pull the live graded version with
`npx super-agent install llm-finetuning-strategist`.