Recommended stack
What are you building? Customise a model
One technology per slot, with the reason it is suggested and what to use instead.
Base model
modelFamily of decoder-only Transformer language models trained with a generative objective.
Why derived from the graph
The agents converse through a model backend, with the hosted models among the supported ones.
Alternatives
Method
techniqueContinue training a pre-trained model on task-specific data to adapt its behaviour.
Alternatives
Efficient alternative
techniqueFine-tuning by training small low-rank adapter matrices instead of all weights.
Why derived from the graph
QLoRA adds 4-bit quantisation of the frozen base model on top of LoRA adapters.
Serving
techniqueCached attention keys and values that make generating one token at a time affordable.
Why derived from the graph
Reusing a shared prefix means reusing the cached keys and values for it.