Updates reddit's description and links.

data/reddit/README.md

@@ -9,11 +9,16 @@ We preprocess the Reddit data released by [pushshift.io](https://files.pushshift
 5. Lowercase the text.
 6. Tokenize the text (using nltk's TweetTokenizer).
 
-We also remove users and comments that simple heuristics or preliminary inspections mark as bots; and remove users with less than 5 or more than 1000 comments (which account for less than 0.01% of users). We include the code for this preprocessing in the ```source``` folder for reference, but host the preprocessed dataset [here](https://drive.google.com/file/d/1CXufUKXNpR7Pn8gUbIerZ1-qHz1KatHH/view?usp=sharing). We further preprocess the data to make it ready for our reference model (by splitting it into train/val/test sets and by creating sequences of 10 tokens for the LSTM) [here](https://drive.google.com/file/d/1lT1Z0N1weG-oA2PgC1Jak_WQ6h3bu7V_/view?usp=sharing). The vocabulary of the 10 thousand most common tokens in the data can be found [here](https://drive.google.com/file/d/1I-CRlfAeiriLmAyICrmlpPE5zWJX4TOY/view?usp=sharing).
+We also remove users and comments that simple heuristics or preliminary inspections mark as bots; and remove users with less than 5 or more than 1000 comments (which account for less than 0.01% of users). We include the code for this preprocessing in the ```source``` folder for reference, but host the preprocessed dataset [here](https://drive.google.com/file/d/1lT1Z0N1weG-oA2PgC1Jak_WQ6h3bu7V_/view?usp=sharing). The vocabulary of the 10 thousand most common tokens in the data can be found [here](https://drive.google.com/file/d/1I-CRlfAeiriLmAyICrmlpPE5zWJX4TOY/view?usp=sharing).
 
 ## Setup Instructions
 
-1. To use our reference model, download the data [here](https://drive.google.com/file/d/1PwBpAEMYKNpnv64cQ2TIQfSc_vPbq3OQ/view?usp=sharing) into a ```data``` subfolder in this directory. This is a subsampled version of the complete data. Our reference implementation doesn't yet support training on the [complete dataset](https://drive.google.com/file/d/1lT1Z0N1weG-oA2PgC1Jak_WQ6h3bu7V_/view?usp=sharing), as it loads all given clients into memory.
+1. To use our reference model, download the data [here](https://drive.google.com/file/d/1ISzp69JmaIJqBpQCX-JJ8-kVyUns8M7o/view?usp=sharing) into a ```data``` subfolder in this directory. This is a subsampled version of the complete data. Our reference implementation doesn't yet support training on the [complete dataset](https://drive.google.com/file/d/1lT1Z0N1weG-oA2PgC1Jak_WQ6h3bu7V_/view?usp=sharing), as it loads all given clients into memory.
 2. With the data in the appropriate directory, run build the training vocabulary by running ```python build_vocab.py --data-dir ./data/train --target-dir vocab```. 
 3. With the data and the training vocabulary, you can now run our reference implementation in the ```models``` directory using a command as the following:
   - ```python3 main.py -dataset reddit -model stacked_lstm --eval-every 10 --num-rounds 100 --clients-per-round 10 --batch-size 5 -lr 5.65 --metrics-name reddit_experiment```
+
+## Comments
+
+- The results in our manuscript were obtained with a previous version of the dataset (see the corresponding [issue](https://github.com/TalwalkarLab/leaf/issues/33)).
+- For an alternative dataset that has been widely benchmarked for the community (more than Reddit), we recommend checking out the [StackOverflow](https://www.tensorflow.org/federated/api_docs/python/tff/simulation/datasets/stackoverflow) dataset.