Bonito: Generate synthetic instruction-tuning datasets

Understand the project

Practical overview

Use Bonito when you have unannotated text and want to generate training examples for one of its supported task types through a Python workflow.

Turn unannotated documents into synthetic examples for a chosen NLP task.
Create question-answering or question-generation data from domain text.
Generate examples for summarization, classification, paraphrase, text generation or natural language inference.

Before you choose

The repository is public and identifies Bonito as open source under the BSD-3-Clause license. The README documents installation with pip.

The input should be unannotated text organized so the relevant context column can be passed to generate_tasks.
The task must match a supported full name or short form; the README lists 16 task types.
The documented workflow uses vLLM sampling parameters alongside the Bonito model and Hugging Face Datasets.
The project links tutorials for a quantized model on a Colab T4 instance and for the model on a Colab A100 instance.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Use Bonito’s documented Python workflow to generate synthetic natural language inference training examples from unannotated text.

Python environment with the bonito-llm package installed.
The Bonito model and dependencies used by the example: Hugging Face Datasets and vLLM.
A dataset with unannotated text in the column supplied as context_col.

Install the package with the command below.
Documented command
```
pip3 install bonito-llm
```
Import Bonito, SamplingParams and load_dataset from their documented packages.
Initialize Bonito("BatsResearch/bonito-v1") and load unannotated text with Hugging Face Datasets.
Set sampling parameters and call generate_tasks with the text, its context column and task_type="nli".

The documented code assigns the generated result to `synthetic_dataset` for natural language inference.

The README’s example uses a particular experiment dataset and selects ten rows; it does not establish performance or suitability for other datasets.
The README links separate Colab tutorials for quantized T4 use and A100 use; the example does not specify general hardware requirements.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 6 November 2026
Original website title: BatsResearch/bonito · Official homepage title · retained in its source language