Model Fine-Tuning

For when prompting isn't enough: we adapt open models to your domain, output format and tone, and measure whether it worked.

Model Fine-Tuning

For when prompting isn't enough: we adapt open models to your domain, output format and tone, and measure whether it worked.

When It's Worth It

We try prompting and retrieval first, because they're cheaper and faster to change. Fine-tuning is worth it in a few specific situations.

Good reasons to fine-tune:

You need a strict, consistent output format at high volume
Your domain has vocabulary or conventions general models get wrong
You want a small, cheap model to match a large one on a narrow task
The model has to run on your own hardware

LoRA, QLoRA & Beyond

Parameter-efficient fine-tuning trains a small adapter on top of an open model. It's fast to train, cheap to store and easy to swap.

Methods:

LoRA and QLoRA on Llama, Qwen, Mistral and Gemma families
Full fine-tuning when the adapter isn't enough
Preference tuning (DPO) for tone and style
Distillation from a larger model into a smaller one

python

1from peft import LoraConfig, get_peft_model
2
3config = LoraConfig(
4    r=16,
5    lora_alpha=32,
6    lora_dropout=0.05,
7    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
8    task_type="CAUSAL_LM",
9)
10model = get_peft_model(base_model, config)
11model.print_trainable_parameters()  # typically well under 1%

Training Data

The quality of the training data does more for the result than the choice of method. Most of a fine-tuning project is spent here.

Data work:

Examples curated from your own data, with duplicates and near-duplicates removed
PII removed before training
Expert annotation and review
Synthetic examples used only where a person has reviewed them

Evaluation

Every fine-tune is compared against the base model and against simply prompting a stronger model. If neither comparison favors it, we tell you.

What we measure:

Task accuracy on a held-out test set built from your real data
Regression checks on general capability
Latency and cost per request compared with the alternatives

Model Fine-Tuning

When It's Worth It

LoRA, QLoRA & Beyond

Training Data

Evaluation

Related capabilities

AI Inference

Private LLM Deployment

Data Science & Machine Learning