A clean training dataset
Examples gathered from your data, de-duplicated, labelled and split, with the edge cases a model most needs to learn.
Fine-tuning is worth it when a base model keeps missing your format, your vocabulary or your language. I build the dataset, train the model and, most importantly, build the evaluation that shows whether it was worth it.
For teams whose prompts have grown into essays, who need lower cost per call, or who work in a domain or language the big models handle poorly.
01/What you get
A tuned model
LoRA or QLoRA on open-weight models (Llama, Qwen, Mistral) or hosted fine-tuning on OpenAI, Vertex AI or Bedrock.
An evaluation suite
Held-out tests, task metrics and model-graded checks comparing tuned against base, so the decision to ship is made on evidence.
Serving and cost model
Deployment on your cloud or a managed endpoint, with a clear per-request cost compared to the model you replace.
02/How it works
- Step 1
Baseline first
Measure the best base model with good prompting. Sometimes that is the answer and we stop there.
- Step 2
Build the data
Curate a few hundred to a few thousand high-quality examples from your real work.
- Step 3
Train and compare
Train small and fast, compare on the held-out set, then scale only what wins.
- Step 4
Deploy and monitor
Ship behind a flag, watch quality in production and keep the dataset growing.
03/Proof
04/Questions
How much data do we need?+
Less than most people think for style and format: a few hundred excellent examples can be enough. Teaching genuinely new knowledge is better done with retrieval.
Will our data be used to train someone else's model?+
No. Training runs in your cloud account or on providers whose terms exclude your data from their training, and you own the resulting weights or adapters.
Do you also evaluate models we already use?+
Yes. Part of my work has been evaluating AI-generated code and model output for training pipelines, and an independent evaluation of an existing system is a common first engagement.
Related writing