Fine-Tuning Services
Make the model yours. Literally.
Fine-tuning turns a general open-weight model into a specialist that speaks your domain, follows your formats, and runs on your hardware. NetRay scopes it honestly (most projects need LoRA and quantization, not a GPU cluster), trains on your infrastructure, and hands you the weights.
The six ways to make a model yours
Fine-tuning is a menu, not a monolith, and vendors love selling the expensive end of it. The honest breakdown:
- Continued pretraining: for genuinely new domains with large private corpora. Rarely needed; expensive.
- Full supervised fine-tuning (SFT): maximal adaptation when you have strong instruction data and the hardware.
- LoRA / QLoRA: 90% of the benefit at a fraction of the cost; trains on a single GPU node. Where most projects should start.
- DPO / preference tuning: teaches the model your judgment calls (tone, safety boundaries, format preferences).
- Distillation: compress a large teacher's capability into a small, fast, cheap-to-serve student.
- Quantization: not training, but the step that makes the tuned model fit your hardware budget.
Data is the project; training is the easy week
The real work is building the training set: mining your tickets, documents, transcripts, and ERP records into instruction pairs; deduplicating and cleaning; and, critically, holding out an evaluation set that reflects the job. We build evaluation first, because a fine-tune you cannot measure is a fine-tune you cannot trust.
All of it can happen inside your perimeter. For regulated clients, training data never leaves the building: we bring the pipeline to the data, run training on your GPUs (or a cluster we stand up for you), and the finished adapter weights stay in your registry.
What a typical engagement looks like
Weeks 1-2: evaluation set construction and baseline benchmarking of 2 to 3 candidate base models on your tasks. Weeks 3-5: data pipeline and first LoRA training runs, iterating against the eval set. Week 6: quantization, serving integration, and a side-by-side report: base model vs tuned model vs (where relevant) the hosted frontier model, on your tasks, with numbers.
The deliverable is not a demo. It is weights in your registry, a reproducible training pipeline, the evaluation harness, and documentation your team can rerun without us.
When NOT to fine-tune
If retrieval solves it, fine-tuning is the wrong tool: facts that change weekly belong in RAG, not in weights. If prompting solves it, ship the prompt. We tell clients this in the first call because a fine-tuning vendor who never says 'you don't need fine-tuning' is selling GPUs, not outcomes. The usual right answer is a small tuned model for form and domain language, plus retrieval for facts.
Straight answers
How much data do we need to fine-tune?
Less than most teams fear. Meaningful LoRA results often start at 1,000 to 10,000 high-quality instruction pairs for a focused task. Quality and coverage beat raw volume; ten thousand clean examples outperform a million noisy ones.
What does fine-tuning cost?
LoRA projects are dominated by data engineering, not compute: single-GPU training runs cost tens of dollars in electricity, and the engagement cost sits in the weeks of data and evaluation work. Full SFT and continued pretraining scale up from there. Bad scoping is the expensive part, which is why we scope with an evaluation set before quoting the full project.
Will a fine-tuned small model really beat a frontier model?
On narrow, well-defined tasks with good training data: frequently, yes, and it serves at a fraction of the latency and cost, on your hardware. On open-ended general reasoning: no. Our benchmarking phase answers this for your specific task before you invest.
Who owns the resulting model?
You do. Weights, adapters, training pipeline, evaluation sets: all delivered into your registry under terms that survive the engagement. Fine-tuned-on-your-data means the weights are part of your IP, and we put that in writing.
Scope the fine-tune with evidence, not vibes.
Bring one task and some example data. We'll build a small evaluation set and tell you honestly whether fine-tuning, RAG, or a better prompt wins.
Scope a fine-tuning projectFree AI workshops available. A founder reads every message.