Local AI Glossary

Small file of trained weights (low-rank matrices) added to the base model without modifying it: a few MB in size, loadable and removable at will, so the same model can carry several adapters for different behaviors.

Additional training process on a pre-trained model, using a specific dataset to specialize it for a particular task. For example, you can fine-tune Llama 7B with legal documents to create a model expert in law. Popular techniques include LoRA and QLoRA.

Low-Rank Adaptation. An efficient fine-tuning technique that updates only a small fraction of the model weights instead of all parameters. LoRA enables fine-tuning a 7B model on a single consumer GPU, generating adaptation files of a few MB instead of gigabytes.

Parameter-Efficient Fine-Tuning. Family of techniques that train a small fraction of the model's parameters (LoRA, adapters, learnable prompts) instead of all weights, cutting training memory and time.

A reinforcement-learning algorithm that optimizes an agent's policy by limiting the size of each update: stable training even with noisy simulated environments. It is the de facto standard for agents that must play, move, or manipulate objects.

Quantized Low-Rank Adaptation. A LoRA variant that quantizes the base model to 4 bits (NF4) before training: it lets you adapt models up to 65B on a single 48 GB GPU, with adapters of a few MB.

Training by trial and reward: the agent acts in an environment and receives a signal (reward) for each outcome. Compared to fine-tuning on hand-written examples, RL learns from complete interactions — and shines when a verifiable task exists, one that says by itself whether it succeeded (e.g. the calculator shows the result). It is the post-training step that took OpenMobile-2 from 31.9 to 59.6 on SPA-Bench.