Update README.md

README.md

@@ -10,6 +10,7 @@ datasets:
 Tk-Instruct is a series of encoder-decoder Transformer models that are trained to solve various NLP tasks by following in-context instructions (plain language task definitions, k-shot examples, explanations, etc). Built upon the pre-trained [T5 models](https://arxiv.org/abs/1910.10683), they are fine-tuned on a large number of tasks & instructions that are collected in the [Natural Instructions benchmark](https://github.com/allenai/natural-instructions), which contains 1600+ tasks in 70+ broach categories in total. This enables the model to not only process the training tasks, but also generalize to many unseen tasks without further parameter update.
 
 More resources for using the model:
+- **Paper**: [link](https://arxiv.org/abs/2204.07705)
 - **Code repository**: [Tk-Instruct](https://github.com/yizhongw/Tk-Instruct)
 - **Official Website**: [Natural Instructions](https://instructions.apps.allenai.org/)
 - **All released models**: [allenai/tk-instruct](https://huggingface.co/models?search=allenai/tk-instruct)
@@ -52,7 +53,9 @@ If you find serious issues or any interesting result, you are welcome to share w
 
 ## Training data
 
-Tk-Instruct is trained using the tasks & instructions in [Natural Instructions benchmark](https://github.com/allenai/natural-instructions), which contains 1600+ tasks in 70+ broach categories in total. We follow the official train/test split. Tk-Instruct model series were trained using 757 tasks, and mTk-Instruct series were trained using 1271 tasks (including some non-English tasks). The training tasks are in 64 broad categories, such as text categorization / question answering / sentiment analysis / summarization / grammar error detection / dialogue generation / etc. The other 12 categories are selected for evaluation.
+Tk-Instruct is trained using the tasks & instructions in [Natural Instructions benchmark](https://github.com/allenai/natural-instructions), which contains 1600+ tasks in 70+ broach categories in total. We follow the official train/test split. Tk-Instruct model series were trained using 757 tasks, and mTk-Instruct series were trained using 1271 tasks (including some non-English tasks). 
+
+The training tasks are in 64 broad categories, such as text categorization / question answering / sentiment analysis / summarization / grammar error detection / dialogue generation / etc. The other 12 categories are selected for evaluation.
 
 
 ## Training procedure
@@ -73,6 +76,4 @@ Although they are trained with only one type of instruction encodings, we found
   eprint={2204.07705},
   primaryClass={cs.CL},
 }
-```
-
-
+```