UniLM
Microsoft's collection of research models and code for language, vision, speech and document AI.
- Stars
- 22.2k
- Forks
- 2.7k
- Licence
- MIT
- Language
- Python
- Last update
- 2 months ago
- Type
- Other
What it is
unilm is a large research repository from Microsoft's GeneralAI group that gathers many pre-trained AI models and their code under one roof. The models cover text in 100+ languages, images, speech, documents with layout, and combined multimodal inputs such as BEiT-3 and Kosmos. It is mainly a research resource, and the README is largely a directory of named projects, each linking to its own subfolder.
- Browse the BEiT, LayoutLM, or TrOCR folders to find a pre-trained model for a specific task
- Use MiniLM-style small models as a starting point for fine-tuning on a language task
- Read the Kosmos or BEiT-3 entries to understand current multimodal research directions
Great for
- Researchers comparing pre-trained model families such as MiniLM, BEiT, or LayoutLM
- Document AI developers looking at LayoutLM, LayoutXLM, or TrOCR
- Teams needing small pre-trained language models like MiniLM for limited hardware
- People exploring multilingual models such as InfoXLM or DeltaLM
Not ideal for
- Anyone wanting a single ready-to-use app; the README is a catalog of separate research projects, not one product
- Non-technical users, since most entries assume machine learning knowledge
- Those needing a single install path, because each subproject has its own setup described in its own folder
How to get it
- 1
Clone and build from source
Clone the repository
git clone https://github.com/microsoft/unilm.git - 2
Run it
- Clone the repository and open the subfolder for the project you want, such as minilm, beit, or layoutlm.
- Read that subfolder's own README for its dependencies and setup steps.
- Install the listed dependencies and follow the example or fine-tuning instructions in that folder.
- 3
Get the most out of it
- Start from the project closest to your task (for example, LayoutLM for documents or TrOCR for OCR).
- Check the News section for recent model releases and their links.
- Use the linked arXiv papers to understand each model before fine-tuning it.
Downloads
Strengths
- Covers many modalities (text, vision, speech, documents) in one organized repository
- Includes well-known research lines with linked papers
- Released under the MIT license
Trade-offs
- The top-level README does not give one install or quick-start path; setup depends on each subproject
- Some entries are marked NEW or are research prototypes whose maturity is not stated here
- Many named projects can be hard to navigate for newcomers
How it compares
| Project | Stars | Type | Level | Updated | Best for |
|---|---|---|---|---|---|
| unilm (this) | 22.2k | Other | advanced | 2 months ago | Researchers comparing pre-trained model families such as MiniLM, BEiT, or LayoutLM |
| AI-For-Beginners | 70.5k | Learning resource | beginner | 13 days ago | Beginners who want a structured, self-paced introduction to AI |
| storm | 31.8k | AI model / agent | intermediate | 1.1 years ago | Developers who want to build a customizable research or report-writing pipeline |
| FinGPT | 21.5k | AI model / agent | advanced | 12 days ago | Developers who want to experiment with finance-tuned LLMs for sentiment analysis |
| nlp-tutorial | 14.9k | Learning resource | intermediate | 2.7 years ago | Students who want to see how classic NLP models work in code |
| Transformers-Tutorials | 11.8k | Learning resource | intermediate | 2 months ago | Developers learning how to use Hugging Face Transformers models in PyTorch |
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.