Update README.md
Browse files
README.md
CHANGED
|
@@ -14,6 +14,28 @@ pipeline_tag: visual-document-retrieval
|
|
| 14 |
|
| 15 |
|
| 16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
## Model
|
| 18 |
This is the model card for `ModernVBERT-embed` the dense encoder version of ModernVBERT not specialised on any tasks, made for general image encoding tasks.
|
| 19 |
|
|
|
|
| 14 |
|
| 15 |
|
| 16 |
|
| 17 |
+
> [!WARNING]
|
| 18 |
+
>🚨 **Migration guide to official `transformers` modeling**
|
| 19 |
+
>
|
| 20 |
+
>ModernVBERT is part of `transformers` since v5.3.0!
|
| 21 |
+
>
|
| 22 |
+
>This becomes the official modeling for the model.
|
| 23 |
+
>
|
| 24 |
+
>If you have a model based on the initial modeling, you can adapt the weigths to the new format supported by `transformers`.
|
| 25 |
+
>We provide a conversion script in this repo, you simply need to run:
|
| 26 |
+
>
|
| 27 |
+
>```bash
|
| 28 |
+
>python convert_model_weights.py \
|
| 29 |
+
> /path/to/legacy-modernvbert \
|
| 30 |
+
> /path/to/converted-modernvbert \
|
| 31 |
+
>```
|
| 32 |
+
>
|
| 33 |
+
>The converted model:
|
| 34 |
+
>- merges the split token embedding tables
|
| 35 |
+
>- rewrites `model.connector.modality_projection.proj.weight`
|
| 36 |
+
>- nests `model.vision_model...` under `model.vision_model.vision_model...`
|
| 37 |
+
|
| 38 |
+
|
| 39 |
## Model
|
| 40 |
This is the model card for `ModernVBERT-embed` the dense encoder version of ModernVBERT not specialised on any tasks, made for general image encoding tasks.
|
| 41 |
|