Make the README more informative with a focus on the project itself by Ingvarstep · Pull Request #347 · urchade/GLiNER

Ingvarstep · 2026-04-20T10:37:42Z

Because the current README highlights other commercial tools and projects, it creates the following issues:

The README doesn't inform well about the key features and benefits of the GLiNER project;
It creates questions about the maintainability of the project and its open-source goals;
The README doesn't clearly describe the contribution guidelines.

I have crafted a new README proposal to address the issues described above and to make it more informative and better capture the GLiNER project itself.

To highlight GLiNER2 and other amazing projects from the Ecosystem, I created a dedicated section listing them.

Looking forward to your suggestions and feedback.

@tomaarsen, it would be awesome to get your review and feedback as well.

Best,
Ihor

tomaarsen

Sure, I can give a review 🤗

Overall some nice extensions I think, and a pretty logical structure, i.e.: title, tags, relevant banner, one paragraph, and then straight into benefits followed by a very solid quick start. I would probably clean up the em dashes though. In my eyes, they read as low-effort, when clearly a lot of love has gone into this project.

Edit: I would more clearly link to the existing models, e.g.: https://huggingface.co/models?library=gliner
I realise now that I miss that a bit.

tomaarsen · 2026-04-20T10:39:26Z

Ooh, looks very clean

It looks minimalistic, but it was actually hard to find the right prompt to generate it that way ;)

tomaarsen · 2026-04-20T10:43:53Z

- [Example Notebooks](https://github.com/urchade/GLiNER/tree/main/examples)
- Finetune on Colab &nbsp;[<img align="center" src="https://colab.research.google.com/assets/colab-badge.svg" />](https://colab.research.google.com/drive/1HNKd74cmfS9tGvWrKeIjSxBt01QQS7bq?usp=sharing)
-## 🛠 Installation & Usage
+---


This is very minor, but I think the --- lines are not needed when you also have a new header starting below each

Actually, I looked into it a bit more, and we can definitely remove the --- lines when there is a new header after them.

tomaarsen · 2026-04-20T10:46:14Z

-## 🛠 Installation & Usage
+---
+
+## Quick Start


I like the updated README here, very bite-sized. All that's missing (I think) is a link to docs below each (sub)section, e.g. below the Output:, the Quanitzation and Compilation, the Serving, etc.). You already have some links below Training, that's good.

tomaarsen · 2026-04-20T10:47:09Z

-Or apply after loading:
+## Serving
+
+GLiNER provides a built-in serving interface for batch inference:


I would add 1 extra sentence to explain why someone would serve instead of using the script above.

Btw, is this merged? I see #346 which seems to refer to this

It's still under review, but API and documentation should be stable.

tomaarsen · 2026-04-20T10:48:00Z

+    quantize=True,
    compile_torch_model=True,
 )
 ```


I'm missing the quantization/compilation gains a bit. As a user I'd like to know what this does for me.

tomaarsen · 2026-04-20T10:49:00Z

+
+---
+
+<div align="center">


Not very crucial in my opinion, but it's fine to keep.

tomaarsen · 2026-04-20T10:51:10Z

- `quantize="bf16"` — bfloat16. Better numerical stability, slightly less speedup (~1.2x).
- `quantize="int8"` — int8 quantization. On CPU, uses built-in FBGEMM int8 kernels (~1.6x speedup). On GPU, uses [torchao](https://github.com/pytorch/ao) int8 weight-only quantization (~50% memory reduction, no speed gain). Intended for models fine-tuned with quantization-aware training (QAT). Stock DeBERTa-based models lose accuracy with int8.
- On CPU, fp16/bf16 quantization reduces memory usage but does not improve speed.
+| Architecture | Description |


Nice to see this, I honestly lost track and I've been following GLiNER-related developments pretty close

tomaarsen · 2026-04-20T10:51:33Z

+| **Uni-encoder** | Strong zero-shot capabilities, supports up to ~50 entity types. The original GLiNER architecture. |
+| **Bi-encoder** | Scalable to massive numbers of entity types via separate text and label encoding. |
+| **RelEx** | Joint NER and relation extraction in a single model. |
+| **GLiNER Decoder** | Hybrid architecture for open NER — entity types are generated with a small decoder for maximum flexibility. |


Suggested change

| **GLiNER Decoder** | Hybrid architecture for open NER — entity types are generated with a small decoder for maximum flexibility. |

| **GLiNER Decoder** | Hybrid architecture for open NER: entity types are generated with a small decoder for maximum flexibility. |

Personal preference: I would remove the em dashes throughout

tomaarsen · 2026-04-20T10:52:12Z

+
+---
+
+## Popular Use Cases


Even better if you can immediately link from a use case to a matching example like these: https://urchade.github.io/GLiNER/usage.html#example-1-extract-company-information

This is a good point, I am adding cross-links to the documentation with real examples.

Co-authored-by: Tom Aarsen <37621491+tomaarsen@users.noreply.github.com>

Ingvarstep · 2026-04-23T14:58:00Z

Hi @urchade , any thoughts on the readme update?

Ingvarstep added 3 commits April 20, 2026 13:11

Make the README more informative with a focus on the project itself

32e319c

update banner background color

193ad47

add contribution guidelines

3f6dd3e

Ingvarstep self-assigned this Apr 20, 2026

Ingvarstep requested review from tomaarsen and urchade April 20, 2026 10:40

tomaarsen reviewed Apr 20, 2026

View reviewed changes

Ingvarstep and others added 6 commits April 20, 2026 14:18

Apply suggestion from @tomaarsen

127259b

Co-authored-by: Tom Aarsen <37621491+tomaarsen@users.noreply.github.com>

Implement Tom's suggestions

175a53f

Merge branch 'new_readme' of github.com:urchade/GLiNER into new_readme

c957b75

update information on serving the model

a05a376

update docs with more usage examples

ebdff94

Merge branch 'main' into new_readme

ee6af51

	\| GLiNER Decoder \| Hybrid architecture for open NER — entity types are generated with a small decoder for maximum flexibility. \|
	\| GLiNER Decoder \| Hybrid architecture for open NER: entity types are generated with a small decoder for maximum flexibility. \|

Conversation

Ingvarstep commented Apr 20, 2026

Uh oh!

tomaarsen left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Ingvarstep commented Apr 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

tomaarsen left a comment •

edited

Loading