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

From a source checkout:

python -m venv .venv
source .venv/bin/activate
python -m pip install .

When mindlab-toolkit==0.2.0 is available on your configured package index, the equivalent package installation is:

python -m pip install 'mindlab-toolkit==0.2.0'

MinT pins and installs its validated SDK dependency automatically:

mindlab-toolkit==0.2.0
tinker==0.22.0

Do not use pip install --no-deps. That option bypasses dependency installation. import mint checks the installed Tinker distribution before importing Tinker internals and reports how to install tinker==0.22.0 when it is missing or incompatible.

The Python package version (0.2.0) and this repository's release tags (for example v2.6.5) are currently separate version identifiers.

Usage

Global endpoint, which is also the default:

export MINT_API_KEY=sk-...
export MINT_BASE_URL=https://mint.macaron.im/train

China endpoint:

export MINT_API_KEY=sk-...
export MINT_BASE_URL=https://mintcn.macaron.xin/train

Then create clients through MinT's Tinker-compatible surface:

import mint

service_client = mint.ServiceClient()

A non-empty MINT_API_KEY takes precedence over TINKER_API_KEY for the process. A non-empty MINT_BASE_URL similarly takes precedence over TINKER_BASE_URL; an explicit client constructor base_url remains the final override. Both supported deployment URLs retain their /train prefix.

Top-level import mint exposes the public Tinker 0.22.0 API directly and mirrors it 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 keeps Tinker's original behavior: tml- and eyJ... are accepted, while sk- is rejected
• MinT keys start with sk-
• import mint validates Tinker 0.22.0 and patches its confirmed standard client construction paths for MinT keys
• Tinker's original ApiKeyAuthProvider class remains unchanged and still rejects sk- when called directly

If you must keep the exact import tinker statement, import mint earlier in the same process before constructing Tinker clients. Importing only tinker does not load MinT compatibility.

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.