Unsloth Zoo

2026.4.6 · active · verified Sun Apr 12

Unsloth Zoo provides a collection of plug-and-play utilities and example modules designed to work with the Unsloth library for efficient LLM fine-tuning. It includes components for dataset loading, SFT (Supervised Fine-Tuning) trainers, tokenizer utilities, and metrics. As of version 2026.4.6, it aims to simplify and standardize common LLM fine-tuning workflows built on Unsloth, with frequent updates aligning with Unsloth's development.

Warnings

breaking Unsloth Zoo's API and underlying logic are tightly coupled with the `unsloth` library. Major and even minor version updates of `unsloth` can introduce breaking changes that require `unsloth-zoo` to be updated simultaneously. Mismatched versions can lead to runtime errors or incorrect behavior.
Fix: Always install `unsloth-zoo` and `unsloth` together, preferably ensuring they are compatible. Refer to the `unsloth-zoo` README or `unsloth` documentation for recommended version pairings. Frequently update both libraries to their latest compatible versions.
gotcha Dataset templates provided by `unsloth_zoo.data` expect specific JSON or text formats. When using custom datasets, ensure they strictly adhere to these templates (e.g., 'instruction', 'input', 'output' fields for chat templates) to avoid parsing errors or incorrect model training.
Fix: Review the source code or documentation for the specific `DatasetConfig` template you are using to understand the expected format. Validate your custom dataset's structure before loading it with `unsloth_zoo.data.loader`.
gotcha Performance and resource usage of `unsloth-zoo` modules are heavily dependent on the `unsloth` backend and correct GPU setup. Misconfigurations (e.g., incorrect CUDA version, missing `bitsandbytes` installation, incompatible PyTorch) will result in poor performance or errors.
Fix: Ensure your system meets `unsloth`'s GPU requirements. Verify `bitsandbytes` is installed for quantization. Follow `unsloth`'s official installation instructions for your specific CUDA/PyTorch setup before attempting to use `unsloth-zoo`.

Install

pip install unsloth-zoo Install Unsloth Zoo

Imports

Trainer

from unsloth_zoo.sft.trainer import Trainer

load_dataset

from unsloth_zoo.data.loader import load_dataset

DatasetConfig
```
from unsloth_zoo.data.configs import DatasetConfig
```
DatasetConfig is nested within the 'configs' submodule, not directly under 'data'.

Quickstart

This quickstart demonstrates how to load a dataset using Unsloth Zoo's data loading utilities. It configures a dataset using a `DatasetConfig` and then loads it with `load_dataset`, showcasing a common first step in preparing data for fine-tuning.

from unsloth_zoo.data.configs import DatasetConfig
from unsloth_zoo.data.loader import load_dataset

# Configure a dataset, e.g., using a predefined template
config = DatasetConfig(dataset_name="alpaca_template")

# Load the dataset
dataset = load_dataset(config)

print(f"Loaded dataset with {len(dataset)} examples.")
if len(dataset) > 0:
    print("First example:")
    print(dataset[0])

view raw JSON →