Fine-tuning Teaching a model your format and style, not your filing cabinet.
Short answer. Fine-tuning means continuing the training of an existing model on your own examples so that it changes how it behaves: its format, tone, vocabulary or a narrow skill. It is not the best way to teach a model facts. A study comparing the two found that retrieval-augmented generation substantially outperformed unsupervised fine-tuning for adding knowledge. Fine-tuning is cheaper than it sounds: QLoRA fine-tuned a 65-billion-parameter model on a single 48 GB GPU, and the LIMA study taught a model to follow response formats from only 1,000 curated examples. But whatever is fine-tuned in is shared by everyone who uses the model and cannot be access-controlled per person.
01
How it works
You start from a trained model, usually an instruction-tuned one, and train it further on pairs of prompts and ideal answers. The weights shift slightly so that the model produces answers like your examples. Full fine-tuning updates every weight, and Hugging Face’s estimator shows that training with the Adam optimiser needs about four times the memory of the model’s weights. Parameter-efficient methods such as LoRA update a small add-on instead, which is why most office-scale fine-tuning uses them.
The LIMA study found that almost all of a model’s knowledge is learned in pre-training, and that a small set of well-chosen examples is enough to teach it a response style. That is the right mental model: fine-tuning shapes behaviour. It is a poor way to add a filing cabinet of facts.
02
Fine-tuning or retrieval?
What you need
Usually better
Why
Answers from the office’s documents, with sources
Retrieval
Retrieval beat unsupervised fine-tuning for knowledge, and it can show the passage used
Facts that change, such as rates or policies
Retrieval
Update the document, not the model
Different people may see different files
Retrieval
A search can filter by permission; fine-tuned weights are shared by every user
A fixed letter format or output schema
Fine-tuning, after trying a good prompt
LIMA learned response formats from 1,000 curated examples
A narrow, repeated task such as sorting bills
Fine-tuning a small model
Examples teach the pattern; a small model is cheap to run
03
What it costs
Examples: LIMA used 1,000 carefully curated prompts and responses. Preparing good examples is usually the largest cost, in staff time.
Hardware: QLoRA fine-tuned a 65B model on one 48 GB GPU. LoRA reduced trainable parameters by 10,000 times against full fine-tuning of GPT-3 175B.
Testing: every fine-tuned model needs a fixed test set, because tuning can improve one behaviour and damage another.
Maintenance: when the base model is replaced, the fine-tune usually has to be redone.
04
Risks to weigh
Leakage: researchers extracted hundreds of verbatim training sequences, including names, phone numbers and email addresses, from GPT-2, and found larger models more vulnerable. Do not fine-tune on records that not everyone with access should see.
No per-user access control: anyone who can use the tuned model can, in principle, draw out what it learned.
Hard to delete: removing one person’s data from trained weights means retraining. Deleting a LoRA adapter file is easier than undoing full fine-tuning.
Licence: check the base model’s licence. Llama’s requires “Llama” at the start of the name of a distributed model improved with its outputs.
05
Where ANVE stands
ANVE’s document question answering works over the office’s own documents and shows the source passage with each answer, and its private GPTs are assistants built on the office’s own material. The software runs today on a development machine; no sealed unit exists yet.
06
Questions people ask
Do I need to fine-tune a model to use my own documents?
No. Retrieval is usually the better first step: it finds the relevant passages, gives them to the model and can show the source. Fine-tuning mainly changes behaviour, not what the model knows.
How many examples does fine-tuning need?
It depends on the task, so find out by testing. The LIMA study used 1,000 curated examples to teach a response style. Start small, measure on a set of examples held back for testing, and add examples where the model fails.
Can I fine-tune on a normal computer?
Small models, yes, with parameter-efficient methods. QLoRA brought a 65B model onto a single 48 GB GPU; an 8B model needs far less. Training still needs a capable GPU or a lot of patience.
Is fine-tuning the same as training from scratch?
No. Training from scratch builds a model from random numbers over trillions of tokens. Fine-tuning adjusts a finished model with far less data and compute.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
