Update README.md
Browse files
README.md
CHANGED
|
@@ -2,6 +2,16 @@
|
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
### Usage
|
| 7 |
|
|
|
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
|
| 5 |
+
**NLI-Mixer** is an attempt to tackle the Natural Language Inference (NLI) task by mixing multiple datasets together.
|
| 6 |
+
|
| 7 |
+
The approach is simple:
|
| 8 |
+
|
| 9 |
+
1. Combine all available NLI data without any domain-dependent re-balancing or re-weighting.
|
| 10 |
+
|
| 11 |
+
2. Finetune several SOTA transformers of different sizes (20m parameters to 300m parameters) on the combined data.
|
| 12 |
+
3. Evaluate on challenging NLI datasets.
|
| 13 |
+
|
| 14 |
+
This model is a ```deberta-v3-base``` finetuned on 20+ NLI datasets.
|
| 15 |
|
| 16 |
### Usage
|
| 17 |
|