Glossary

Quick reference for core terms in the MinT docs.

Training

Term	Meaning
SFT	Supervised fine-tuning. Train the model to imitate target behavior from labeled input-output pairs.
RL	Reinforcement learning. Optimize the model with reward signals; common algorithms GRPO / PPO.
DPO	Direct preference optimization. Optimize a policy directly from preference pairs, no explicit reward model.
LoRA	Low-rank adaptation. Freeze the base model weights and train only injected low-rank matrices, sharply reducing training and storage cost.
adapter	The artifact produced by LoRA training; mounts onto a base model to change its behavior.
checkpoint	A training snapshot holding adapter weights and optimizer state, used to resume training or sample.
policy	The model being optimized in RL.
rollout	The process of sampling trajectories from a policy during RL.
reward	The scalar signal measuring trajectory quality in RL.

Models & Inference

Term	Meaning
macaron	The model series trained on MinT; current version macaron v1.
base model	The host model an adapter mounts onto.
token	The minimal text unit the model processes; corresponds to integers output by the tokenizer.
sampling	The process of generating text from a model.
logprob	The model's log-probability of a token, used for evaluation and RL.

Platform Components

Term	Meaning
MindForge	MinT's web training console for managing SFT / RL / DPO jobs.
MinT SDK	The client-side Python library; import mint then drives training and sampling.
Model Router	MinT server component that routes inference requests to the right GPU.
Task Manager	MinT server component that schedules training jobs.
LoRA Manager	MinT server component that manages the adapter lifecycle.

Note

This page is for end users.