Mind Lab Toolkit (MinT)

The Open Infrastructure for Experiential Intelligence.

MinT is the reinforcement learning infrastructure for agents and models to learn from real experience. It focuses on the engineering and algorithmic realization of RL across multiple models and tasks, with emphasis on making LoRA RL simple, stable, and efficient.

Visit the MinT website.

Installation

cd mindlab-toolkit
pip install -e .

MinT pins the validated Tinker SDK dependency. If your environment already has a different Tinker version, reinstall with:

python -m pip install --force-reinstall 'tinker==0.15.0'

import mint also checks the installed Tinker version at runtime and fails fast with this command if the version is unsupported.

Usage

import mint

# Set API key via environment variable MINT_API_KEY.
# You can keep both MINT_* and TINKER_* variables in the same .env.
# Importing mint makes MINT_* take precedence for this process; set MINT_BASE_URL
# if you want a non-default endpoint.
# Default base URL: https://mint.macaron.xin
# Mainland China endpoint override: https://mint-cn.macaron.xin/

service_client = mint.ServiceClient()

Top-level import mint is the Tinker-compatible surface. All public Tinker APIs are available directly from mint and mirrored in mint.tinker.

Lowest-Friction Tinker Migration

If your existing code starts with import tinker, the smallest working MinT migration is:

import mint as tinker

Then switch your credentials and endpoint to MinT.

Why this matters:

• raw upstream import tinker still validates API keys with the tml- prefix
• MinT keys start with sk-
• import mint applies the MinT compatibility patches that let the Tinker-style client surface keep working with MinT credentials

If you must keep the exact import tinker statement, import mint earlier in the same process before constructing Tinker clients.

MinT Extension Namespace (mintx)

MinT-only APIs live under mint.mint. The intended usage is:

import mint
import mint.mint as mintx

training_client = ...  # a mint/tinker TrainingClient
mintx.forward_backward_reverse_kl(
    training_client,
    reference_model_path="mint://teacher-step-0010",
    data=[...],
)

Use this namespace for MinT-specific extensions that should not appear in the default top-level mint surface. The current extensions are the MinT-only endpoints forward_backward_reverse_kl and interpolate_checkpoints.

Documentation

Read the MinT documentation at mint-doc.macaron.im.

License

MIT

A Mind Lab Contribution - A Lab for Experiential Intelligence.