Commit
•
2768007
1
Parent(s):
75cd11c
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
datasets: natural_questions
|
| 5 |
+
widget:
|
| 6 |
+
- text: "Who added BigBird to HuggingFace Transformers?"
|
| 7 |
+
context: "BigBird Pegasus just landed! Thanks to Vasudev Gupta, BigBird Pegasus from Google AI is merged into HuggingFace Transformers. Check it out today!!!"
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
This checkpoint is obtained after training `FlaxBigBirdForQuestionAnswering` (with extra pooler head) on [`natural_questions`](https://huggingface.co/datasets/natural_questions) dataset on TPU v3-8. This dataset takes around ~100 GB on disk. But thanks to Cloud TPUs and Jax, each epoch took just 4.5 hours. Script for training can be found here: https://github.com/vasudevgupta7/bigbird
|
| 11 |
+
|
| 12 |
+
**Use this model just like any other model from 🤗Transformers**
|
| 13 |
+
|
| 14 |
+
```python
|
| 15 |
+
from transformers import FlaxBigBirdForQuestionAnswering, BigBirdTokenizerFast
|
| 16 |
+
|
| 17 |
+
model_id = "vasudevgupta/flax-bigbird-natural-questions"
|
| 18 |
+
model = BigBirdForQuestionAnswering.from_pretrained(model_id)
|
| 19 |
+
tokenizer = BigBirdTokenizerFast.from_pretrained(model_id)
|
| 20 |
+
```
|
| 21 |
+
|
| 22 |
+
In case you are interested in predicting category (null, long, short, yes, no) as well, use `FlaxBigBirdForNaturalQuestions` (instead of `FlaxBigBirdForQuestionAnswering`) from my training script.
|