What happens in a fine-tuning project?
Most projects run in the same order. You collect and clean examples, format them into the structure the training tool expects, train, then evaluate against a held-out set and against the untuned model. OpenAI's API documents four types: supervised fine-tuning on input and output pairs, Direct Preference Optimization (DPO) on pairs of better and worse responses, reinforcement fine-tuning that trains a reasoning model against a feedback signal you define, and vision fine-tuning on images with desired outputs.
OpenAI's own cookbook recommends a two-stage path: run supervised fine-tuning on a subset of your preferred responses first, then apply DPO starting from that tuned model. DPO fits subjective tasks where you can say which of two answers is better without being able to write down a single correct answer. Summaries and brand tone are typical examples. DPO also adds its own hyperparameter, beta, which you will tune along with the usual learning settings.
Where do adapters and open models come in?
When a team tunes open-weight models on its own hardware, the work usually runs through Hugging Face libraries. TRL provides an SFTTrainer and a DPOTrainer, plus a command-line tool that can also train a reward model. The PEFT library handles parameter-efficient methods such as LoRA, IA3 and AdaLoRA. LoRA trains a small set of extra weights while the base model stays frozen, so the saved checkpoint holds only the adapter.
Adapters change the engineering. One base model can carry several adapters that you load, switch or hotswap. QLoRA pairs a quantized base model with LoRA to cut memory use. Expect to work on things like rank and alpha settings, training only on the completion part of each example, and distributed training setups such as FSDP. Much of the skill is in debugging data formatting and memory limits, more than in the training call itself.
When is prompting or retrieval the better choice?
Deciding against fine-tuning is part of the job. Databricks frames the choice this way: retrieval-augmented generation (RAG) adds new knowledge at inference time, and fine-tuning is better for changing behavior, tone, format or task structure. If answers must come from current or private documents, you start with RAG. Google Cloud's guidance adds that with only a small amount of data, prompt engineering may be enough.
Fine-tuning earns its cost when you need consistent behavior and fast responses, since RAG adds embedding, search and ranking steps before generation. The two also combine. A tuned model can handle voice and format while a retrieval layer supplies facts that change. Keeping a document index current is ongoing engineering work, and so is retraining a tuned model when the base model or the task changes. Weigh both maintenance costs before you commit.
Who does this work and under what titles?
Fine-tuning rarely has its own job title. It usually sits inside machine learning engineer, applied scientist or research engineer roles, and sometimes inside AI engineer roles at companies building products on hosted model APIs. Model developers and AI labs do post-training at the largest scale. Platform vendors such as OpenAI and Databricks publish fine-tuning tooling and guidance, which tells you their customers do this work too. Product companies with private data and narrow, repeated tasks are the most likely to tune in-house.
Ads for this work typically ask for Python and PyTorch, experience with training libraries like Hugging Face Transformers, TRL and PEFT, solid evaluation design, and the judgment to explain why tuning beats a prompt or a retrieval pipeline. Treat that list as a pattern seen across postings. It is no fixed standard.
Pay
USD a year, before tax. US figures. BLS OEWS May 2025 for software developers, moved to 2026 with the Employment Cost Index; BLS has no AI or fine-tuning series, so this is the nearest occupation and covers all software developers. Job ad rows are midpoints of base pay ranges in 139 live AutomationRoles.ai machine learning engineer ads from 56 employers on 5 November 2026. Those ads lean toward large AI and tech employers and exclude stock and bonus. Neither source isolates fine-tuning specialists. Refer to the Machine Learning Engineers salary guide for figures from live adverts.
What you need
- Python and a deep learning frameworkTraining code, data loaders and evaluation scripts are written in Python, usually with PyTorch underneath libraries like Transformers.
- Dataset constructionYou need to build clean input and output pairs for supervised tuning and ranked response pairs for preference training, in the exact format the tool expects.
- Training methodsKnow supervised fine-tuning, DPO and reward modeling, and when reinforcement fine-tuning with a defined grader fits a task.
- Parameter-efficient tuningBe comfortable with LoRA and QLoRA through PEFT, including rank settings and managing several adapters on one base model.
- EvaluationDesign held-out tests and side-by-side comparisons that show whether the tuned model beats the base model plus a good prompt.
- Architecture judgmentExplain when RAG or prompt engineering solves the problem more cheaply, and when a hybrid of tuning and retrieval is worth it.
How to get there
- Week 1Frame the problemWrite down the target behavior, and try strong prompts and retrieval first to set a baseline.
- Weeks 1-3Build the datasetCollect, clean and format examples, and hold out a test set the model never sees.
- Weeks 2-4Supervised fine-tuningTrain on input and output pairs, often as a LoRA adapter, and compare against the baseline.
- Weeks 4-6Preference trainingIf quality is subjective, gather better and worse response pairs and run DPO on top of the SFT model.
- Weeks 5-7Evaluate and decideRun the held-out tests and side-by-side reviews, then ship, iterate, or drop tuning for prompting or RAG.
- OngoingMaintainRetrain when the base model, data or task changes, and keep monitoring output quality in production.
Open now on AutomationRoles.ai
Questions people ask
Is fine-tuning a separate job from machine learning engineering?
Usually no. It is a set of skills inside machine learning engineer, applied scientist or research engineer roles. Some AI engineer roles include it when a team tunes hosted models through an API.
What is the difference between SFT and DPO?
Supervised fine-tuning trains a model on inputs paired with desired outputs. DPO trains on pairs of responses where one is preferred, which suits tasks with no single correct answer. OpenAI recommends running SFT first and then applying DPO to the tuned model.
Why use LoRA instead of full fine-tuning?
LoRA trains a small number of added weights and keeps the base model frozen. Checkpoints only store the adapter, so they are smaller and faster to save, and one base model can switch between several adapters.
Should I fine-tune or use RAG?
Use RAG when answers depend on new or private documents that change. Fine-tuning fits changes to behavior, tone, format or task structure, and it can lower latency. Many systems combine the two.
Is there a pay figure specific to fine-tuning roles?
No source used here isolates fine-tuning specialists. The nearest guides are the BLS software developers series and the board's machine learning engineer job ads. Both cover wider groups than fine-tuning alone.
Do I need a certification or license to do this work?
No license exists for fine-tuning work, and no regulator oversees the job. Employers mostly judge you on shipped projects and on how well you can explain your evaluation results.
Sources
- OpenAI: Model optimization track
- OpenAI Cookbook: Choosing between SFT, DPO and RFT
- OpenAI: Direct preference optimization guide
- OpenAI: Reinforcement fine-tuning guide
- Hugging Face: TRL quickstart
- Hugging Face: Parameter-efficient fine-tuning in Transformers
- Hugging Face: LoRA and PEFT course unit
- Databricks: RAG vs fine tuning
- Google Cloud: To tune or not to tune
- BLS: Software developers, Occupational Outlook Handbook
