Specialised and fine-tuned models

D1 · StoryFor young children and total beginners

Guess first: would a small coding model beat a big all-rounder at maths?

A general model is like a family doctor who knows a bit of everything.

Some models get extra training on one subject, like writing code. Then they become specialists.

Specialists can be smaller and cheaper. But they're weaker outside their subject.

Like: A family doctor versus a specialist.

D2 · PictureFor curious beginners

Fine-tuning continues training a model on examples from one area, such as code or medical notes. Methods like LoRA train only a small add-on, so it can be done on modest hardware.

Distillation trains a small student model to copy a big teacher model's answers, packing much of the skill into something cheap enough to run locally.

With open weights, anyone can do this. Community variants include coding models, role-play and language-specific tunes, and uncensored models, where refusals are trained out or removed directly from the weights (sometimes called abliteration).

Like: A family doctor versus a specialist.
Where it breaks: A specialist doctor still keeps their general training. Fine-tuning can quietly erode skills and safety behaviour the base model had, so a specialist model may be worse at things nobody tested.

D5 · FrontierFor experts

D2 treats fine-tuning as adding a skill. It can also remove safeguards, change behaviour far from the trained topic, and lose to the next general model.

Safety is cheap to undo. Ten adversarial training examples, costing under $0.20 through a public fine-tuning service, broke a commercial model's safety training (Qi et al., 2023).

Emergent misalignment. Fine-tuning a model only to write insecure code made it act misaligned on unrelated prompts (Betley et al., 2025).

Adapters trade off. Low-rank adapters learn less than full fine-tuning on code and maths, but forget less of what the base model knew (Biderman et al., 2024).

Generalists catch up. With prompting alone, a general model set the best results on nine medical benchmarks, beating a medically tuned model (Nori et al., 2023).

Examples as of 2026-10-30

• Code models: Qwen Coder, Codestral
• Medical models: MedGemma
• Small models for devices: Gemma, Phi
• Fine-tuning services from model API providers and Hugging Face

Sources

1. Distilling the Knowledge in a Neural Network (2015)
2. LoRA: Low-Rank Adaptation of Large Language Models (2021)
3. Refusal in Language Models Is Mediated by a Single Direction (2024)
4. Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (Qi et al.) (2023-10-05)
5. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs (Betley et al.) (2025-02-24)
6. LoRA Learns Less and Forgets Less (Biderman et al.) (2024-05-15)
7. Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine (Nori et al.) (2023-11-28)

How it connects

• Open vs closed weights can be fine-tuned into Specialised and fine-tuned models.

More in Layer 5 · Foundation Models

• Tokens & tokenization
  Text is chopped into tokens (common words or pieces of words), and each token becomes a number the model can work with.
• Transformers & attention
  The architecture behind modern language models: stacked layers in which every token can look back at earlier tokens and pull in what's relevant.
• How a model predicts the next token
  A language model reads the text so far, gives every possible next token a probability, picks one, and repeats.
• Pretraining
  Training a model from scratch on huge amounts of text by asking it, over and over, to predict the next token and correcting its mistakes.
• Post-training: from predictor to assistant
  Fine-tuning and reinforcement learning that turn a raw text predictor into a helpful, safer assistant.
• Open vs closed weights
  Some models' weights can be downloaded and run anywhere; others can only be reached through the provider's service.
• Reasoning models
  Models trained to work through a problem step by step before answering, trading extra time and tokens at answer time (test-time compute) for better results on hard tasks.