Base model
The raw, pretrained model before anyone taught it to follow instructions or be polite — powerful but unruly.
Think of it like
A brilliant improviser who hasn’t learned to answer the actual question you asked.
Example
Hand a base model “Q: capital of France? A:” and it may just keep inventing more Q&As.
How it actually works
A base model only knows how to continue text, because that’s all it was trained to do. The chat model you actually use is a base model plus alignment steps (instruction tuning, preference training) that teach it to treat your message as a request to fulfill rather than a passage to extend.
For product teams
Usually not what you want directly — the instruction-tuned version is the usable product.
For engineers
Pretrained-only checkpoint (next-token objective), pre-alignment; completes rather than instructs.
Related
- Fine-tuning — Turned helpful by instruction tuning / fine-tuning.
- Foundation model — The starting point for a foundation model’s products.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome