Developers

Bring your own table

Install once. Choose a task. Predict.

Pass labeled context and query rows to LimiXPredictor for classification, regression, or missing-value imputation.

INFERENCE PATHNO PARAMETER UPDATE

CONTEXTx_train + y_trainLabeled rows

QUERYx_testRows to resolve

MODELLimiX-2Pretrained checkpoint

OUTPUT

class probabilities
continuous targets
imputed features

Context and query arrays enter the same checkpoint. The selected task configuration determines the output.

Installation and Quick Start

Run your first prediction

Clone the repository, install the package, then run a complete classification example from the repository root.

INSTALL FROM SOURCE

$git clone https://github.com/limix-ldm-ai/LimiX.git$cd LimiX$python -m pip install -e .

PYTHON API

Minimal inference flow

Save as quickstart.py and run it from the cloned repository root.

QUICKSTART.PYPYTHON ≥ 3.12

import os

os.environ.update(
    RANK="0",
    WORLD_SIZE="1",
    MASTER_ADDR="127.0.0.1",
    MASTER_PORT="29500",
)

import numpy as np
import torch
from huggingface_hub import hf_hub_download
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from inference.predictor import LimiXPredictor

x, y = load_breast_cancer(return_X_y=True)
x_train, x_test, y_train, _ = train_test_split(
    x, y, test_size=0.5, random_state=42
)
checkpoint = hf_hub_download(
    repo_id="stable-ai/LimiX-2",
    filename="LimiX-2.ckpt",
    local_dir="./cache",
)
predictor = LimiXPredictor(
    device=torch.device("cuda"),
    model_path=checkpoint,
    inference_config="config/cls_default_noretrieval_v2.json",
)
probabilities = predictor.predict(
    x_train, y_train, x_test, task_type="Classification"
)
print(probabilities[:3])
print(np.argmax(probabilities, axis=1)[:3])

COMMAND LINE

Run classification with
LimiX-infer

Replace both absolute paths, then run the command after the editable install.

TERMINALLIMIX-INFER

LimiX-infer \
  --task_type cls \
  --data_dir /absolute/path/to/classification-benchmark \
  --model_path /absolute/path/to/LimiX-2.ckpt \
  --gpuid 0 \
  --save_name limix_cls

--data_dir: Benchmark root containing one directory per dataset, with matching *_train.csv and *_test.csv files.
--model_path: Absolute path to the downloaded LimiX-2.ckpt.

Task aliases: cls, reg, and imputation.

Read the full CLI guide

INTERFACEProvide a checkpoint and prepared context/query arrays.

LICENSELimiX-2 weights are non-commercial. Commercial use requires a separate agreement; the full license governs use and redistribution.

Interface checked against the official documentation and repository on 20 Sep 2026. Python 3.12 or later is required. Match PyTorch and optional flash-attn to your CUDA environment using the repository guide.

Choose a task

One predictor for many inference tasks

The checkpoint and prediction interface are shared across tasks. Create the predictor with the configuration that matches your task.

Classification