txtinstruct: Build Instruction-Tuned Models from Your Data

Overview

txtinstruct helps developers create instruction-following datasets and use them to train instruction-tuned models. Its focus is on open data and models, with support for bringing your own data, addressing the unclear licensing of some existing datasets and language models.

It is most relevant to practitioners who want to build a dataset-to-model workflow rather than simply use a pre-trained assistant. The project is built on txtai and supports Python 3.8 and later.

Key Features

• Build instruction-following datasets from your own data.
• Train instruction-tuned models using those datasets.
• Support workflows centered on open data and open models.
• Provide a framework for combining dataset creation and model training.
• Include an example notebook that demonstrates building datasets and models.
• Integrate with txtai.

Use Cases

• ML practitioners preparing domain-specific data to train an instruction-tuned model.
• Teams seeking more control over the licensing of training datasets.
• Researchers exploring workflows based on open datasets and models.
• Developers who want to experiment with instruction tuning using Python and txtai.

Project Facts

• Language: Python
• License: Apache-2.0
• Stars: 237
• Forks: 11
• Topics: none listed
• Archived: yes

Getting Started

Install from PyPI:

pip install txtinstruct

See the repository README for the example notebook and further details.

Alternatives

• xTuring: xTuring covers broader model customization and deployment, while txtinstruct focuses on creating instruction-following datasets and training on them.
• LlamaFactory: LlamaFactory offers broad fine-tuning workflows for existing datasets, while txtinstruct centers on creating instruction-following datasets.
• torchtune: torchtune provides configurable LLM post-training recipes, while txtinstruct focuses on instruction dataset creation alongside training.

Considerations

The repository is archived, and its last recorded push was on 2023-09-23, so prospective users should account for the possibility that it is no longer maintained. The project supports Python 3.8 and later and depends on txtai. The supplied information does not specify hardware requirements or provide current compatibility details for model-training environments.