Who this is for
- RAG or prompting already in production, hitting a tone / format / classification wall
- A narrow high-volume task where a smaller tuned model beats a frontier API
- Data you can curate and are allowed to train on
Who this is not for
- “Make it know our PDF wiki” — that is RAG / on-prem, not a fine-tune
- RLHF theater on forty examples
What we deliver
- Use-case analysis: fine-tuning vs RAG vs better prompting — honest recommendation first
- Dataset preparation and synthesis from your docs, tickets, transcripts, and style guides
- LoRA / QLoRA adapter training on local or rented compute, with experiment tracking
- Structured evaluation: task-specific benchmarks before and after, so improvement is measurable
- Deployment of adapters into your existing serving stack (vLLM, llama.cpp, Ollama)
How an engagement runs
- Decide. Fine-tune vs RAG vs prompt. Most teams should not train yet. We will say so.
- Dataset. Curate or synthesize examples you own. Quality beats volume.
- Train + eval. LoRA / QLoRA against a frozen baseline. If it does not beat the base, we do not ship the adapter.
- Deploy. Drop the adapter into the runtime you already run. You own the files.
Stack we actually use
- LoRA / QLoRA
- PEFT
- experiment tracking
- vLLM / llama.cpp / Ollama serving
Tested on our fleet
Dated lab notes — not a client logo wall.
Related services
Common questions
Should we fine-tune or use RAG?
Different tools for different problems. RAG injects fresh facts at query time; fine-tuning changes how a model behaves and sounds. Most businesses need RAG first. Fine-tuning earns its keep for specialized formats, classification, tone, and narrow high-volume tasks. We'll tell you honestly which one your use case needs.
How much training data does fine-tuning require?
Far less than most teams assume — often a few hundred to a few thousand high-quality examples with LoRA methods. Data quality matters much more than quantity, and part of our service is curating and synthesizing that dataset with you.
Who owns the fine-tuned model?
You do. Adapters, datasets, and evaluation results are delivered to your infrastructure. No ongoing royalties, no platform dependency.
LoRA vs a full fine-tune?
LoRA / QLoRA is the default: cheaper, reversible, and enough for tone, format, and narrow tasks. Full fine-tunes are a special case we will not default to.
Do we need GPUs to train, or can we rent?
Either. We train on your box when it fits, or on rented compute when it does not. Inference still lands on infrastructure you control.
What if the adapter does not beat the baseline?
We stop. A worse model does not ship. You keep the eval notes either way.