Mind Lab Toolkit (MinT)
Customize

Customize

This section is the reference for everything you can configure in a MinT training job. It is organized by algorithm (SFT, DPO, RL, VLA) plus cross-cutting concepts and recipes. Every leaf page declares one of three fixed templates so the right-side TOC reads the same way across the section.

Where to start

All parameters by algorithm

The full parameter table for each algorithm lives on its own page, in the ## All Parameters section. Quick links:

Concepts index

Foundational topics that show up across multiple algorithms:

TopicPage
Rendering / tokenizationRendering
Loss functions catalogLoss Functions
Completers (TokenCompleter, MessageCompleter, LLM-as-judge)Completers
Weights / checkpoints / TTLWeights
Evaluations (custom, NLL, Inspect AI)Evaluations
Async patterns / num_samplesAsync Patterns

Recipes index

End-to-end recipes that combine multiple primitives, plus deployment paths:

RecipePage
RLHF 3-stage pipelineRLHF Pipeline
Multi-turn RLMulti-turn
Multi-agent RLMulti-agent
Prompt distillationDistillation
Custom environmentCustom Environment
Export to HFExport to HF
LoRA AdapterLoRA Adapter
Publish to HubPublish to Hub

Tinker compatibility. MinT's client SDK is Tinker-API compatible (pip install from mindlab-toolkit). Code patterns that work for Tinker work for MinT — substitute the endpoint (mint.macaron.xin or mint-cn.macaron.xin) and the MINT_API_KEY. See Human Quickstart → Migrating from Tinker for the full one-line migration.

On this page