Pretraining and fine-tuning

In plain English

Pretraining is the heavy first stage: the model reads enormous text corpora and learns next-token prediction. That pass builds grammar, facts, coding patterns, and a little of almost everything on the public internet—plus whatever licensed data the lab added.

Fine-tuning is the second act—smaller, targeted training that steers the model toward helpful chat, a company tone, tool formats, or safer refusals. You can think of pretraining as general education and fine-tuning as job training.

How it works

Pretraining uses self-supervised objectives (predict the next token, sometimes fill-in-the-blank variants). Fine-tuning may use curated prompt–response pairs (supervised fine-tuning), human or model preference rankings (RLHF-style methods), or lightweight adapters that train only a small set of weights.

Instruction-tuned chat models are still next-token predictors—the tuning data just makes “helpful assistant” completions more likely when users ask questions.

• Pretraining: massive data, weeks/months of compute, base model.
• SFT: examples of desired answers on specific prompts.
• Preference tuning: reward models or ranking losses to align style and safety.
• Optional: domain fine-tunes for medicine, law, or internal docs.

Going deeper

Parameter-efficient fine-tuning (LoRA and friends) updates low-rank matrices instead of the full weight matrix—cheaper for teams customizing open models.

Fine-tuning improves average behavior but does not guarantee factual accuracy on rare facts; evaluation and retrieval still matter for production.

Common misconceptions

Fine-tuning teaches the model new facts perfectly.

It shifts behavior and style; factual updates are better handled with retrieval, tools, or targeted data curation—not a single short fine-tune alone.

Pretraining data is exactly what the model will repeat.

Training mixes, filters, and deduplicates vary by lab; the model generalizes and hallucinates—not a verbatim archive.

Key facts

• Pretraining learns general language statistics from large unlabeled text.
• Fine-tuning uses smaller curated datasets for behavior and format.
• Chat models are typically base models plus instruction and preference training.
• Adapters can specialize a model without full-weight retraining.
• Alignment training reduces some harms but does not eliminate hallucinations.