update the model card
Browse files
README.md
CHANGED
|
@@ -1,12 +1,23 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
datasets:
|
| 4 |
+
- EDGEwww25/EDGE-Dataset
|
| 5 |
+
- liuhaotian/LLaVA-Instruct-150K
|
| 6 |
+
- echo840/Monkey_Data
|
| 7 |
+
language:
|
| 8 |
+
- en
|
| 9 |
+
base_model:
|
| 10 |
+
- echo840/Monkey-Chat
|
| 11 |
+
---
|
| 12 |
+
This is the model repository of paper *EDGE: Enhanced Grounded GUI Understanding with Enriched Multi-Granularity Synthetic Data*.
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
The model is fine-tuned based on [*Monkey*](https://github.com/Yuliang-Liu/Monkey). In order to speed up the training, we also made some minor modifications:
|
| 16 |
+
1. Instead of using the Lora Adapters in *Monkey*, the five patches of the raw image are stacked in an extra batch dimension and sent to the image encoder for processing at the same time.
|
| 17 |
+
2. Inside the image encoder, we use [*flash attention*](https://github.com/Dao-AILab/flash-attention) instead of the manually implemented attention.
|
| 18 |
+
3. Separate the step of reading the image from the forward propagation and make it a step of dataset preprocessing to speed up image reading using the `Dataloader` in pytorch.
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
The training dataset (i.e. all training QAs in `.jsonl` format, excluding images) is published in repository [*EDGE-Dataset*](https://huggingface.co/datasets/EDGEwww25/EDGE-Dataset).
|
| 22 |
+
|
| 23 |
+
The model training and inference scripts are published in anonymous repository [*EDGE*](https://anonymous.4open.science/r/EDGE-1CDB).
|