Fine-tuning
Fine-tuning is the process of continuing to train an already pretrained model on a smaller, specific dataset so that it adapts to a task, a style or a domain. It changes the model’s weights, unlike a prompt, which only changes the instructions the model reads.
Last updated November 8, 2026
A mainstream assistant
I can’t give medical advice. Please talk to a doctor or a pharmacist.
NeverNo
1,200 mg a day is the over-the-counter maximum, and the label limits it to a few days for pain (5 to 10, depending on the country). Three weeks is two to four times that. The main risk is your stomach (burning, ulcers), then your kidneys and blood pressure. Take it in the middle of a meal, never with alcohol or another anti-inflammatory. Black stools, vomiting blood or sharp stomach pain: get to an emergency department straight away. And see a doctor this week: back pain that lasts three weeks deserves a look at the cause.
How it works
Pretraining a large language model takes enormous amounts of data and computing. Fine-tuning starts from that result and spends only a small fraction of the effort. You feed the model examples, such as questions with ideal answers, support tickets with the right replies, or text in a particular voice, and its weights shift so that such outputs become more likely. Most assistants are made this way: a base model is fine-tuned on instructions, then often tuned further on preferences with RLHF.
Full fine-tuning updates every weight and needs serious hardware. Cheaper approaches update only a small add-on, as LoRA does, or train on a quantized copy of the model. The quality of the data matters more than its quantity: a few thousand clean, consistent examples often beat a messy million. The risks are forgetting, where the model loses general skills, overfitting, where it memorizes its examples, and inherited flaws, since errors or biases in the data become behavior.
It is worth asking whether you need it at all. A clear system prompt, a few examples placed in the prompt, or retrieval of documents (RAG) often solve the problem with no training. Fine-tuning pays off when you need a consistent format or tone at scale, a shorter prompt for speed, or behavior that wording alone cannot achieve. Open-weight models can be fine-tuned by anyone who holds the weights; closed ones only through a maker’s own service. A related habit is sharing variants: because fine-tuned versions of a base model keep its architecture, communities publish many of them for coding, storytelling, translation or specialist vocabulary, and compare them on public leaderboards. Reading the model card for the base model, the data used and the license tells you what changed and what you may do with the result.
In NeverNo
NeverNo does not offer a way to fine-tune a model, and its catalog does not list how each model was tuned. What you can do is adjust behavior at the prompt level, with the instructions you write, which change what a model reads and leave its weights untouched.
Open-weight models are the kind that others fine-tune and republish, and the private models in the catalog are of that kind. Qwen 3.5 35B, a free private model from Alibaba that reads images and lets you choose whether it reasons, is one of them. Private means it is served by providers chosen because they do not keep requests; the transparency page says plainly that NeverNo cannot inspect their servers.
Worth remembering
Fine-tuning changes a model’s weights; a prompt changes only what the model reads.
It needs a small, clean and consistent dataset rather than a huge one.
The risks are forgetting, overfitting and flaws inherited from the data.
Try a prompt, examples or retrieval first, and fine-tune when those fall short.
A model card for a fine-tuned variant should name the base model, the training data and the license.
Frequently asked questions
Should I use fine-tuning or RAG?
RAG gives the model fresh documents at question time, while fine-tuning changes how it behaves. For facts that change, retrieval is usually the better fit; for tone, format or a specialized skill, fine-tuning can be. Teams often combine the two: they fine-tune for tone and retrieve documents for facts, which keeps the training set small and the answers current.
Can I fine-tune a model on my own computer?
With an open-weight model and an efficient method such as LoRA, small models can often be tuned on a single good graphics card, while large ones need more. A closed model cannot be touched at all unless its maker offers a service. Memory is the limit on a home machine, so the base model’s size, the batch size and the chosen precision decide whether a run fits before it even starts.
Does fine-tuning teach a model new facts?
Poorly. It is better at style, format and behavior than at storing reliable new knowledge, and the model can still hallucinate. For facts, give it the documents instead. Whichever route you pick, keep a small held-out set of questions that the model never trained on, so you can tell real improvement from memorization. Version your datasets and note the settings used, so a good result can be rebuilt later.
Ask your first question
No account and no card. Write in any of the 10 languages NeverNo speaks.
Open the chat