Update README.md

README.md

@@ -8,7 +8,7 @@ widget:
 - text: "World War 2| followed by"
   example_title: "followed by"
 ---
-This is a t5-small model trained from scratch on WikiKG90Mv2 dataset. Please see https://github.com/apoorvumang/transformer-kgc/ for more details on the method.
+This is a t5-small model trained from scratch on WikiKG90Mv2 dataset. Please see https://github.com/apoorvumang/transformer-kgc/ for more details on the method. 
 
 This model was trained on the tail entity prediction task ie. given subject entity and relation, predict the object entity. Input should be provided in the form of "\<entity text\>| \<relation text\>".
 
@@ -16,7 +16,7 @@ We used the raw text title and descriptions to get entity and relation textual r
 
 We trained the model on WikiKG90Mv2 for approx 1.5 epochs on 4x1080Ti GPUs. The training time for 1 epoch was approx 5.5 days.
 
-To evaluate the model, we sample 300 times from the decoder for each input (s,r) pair. We then remove predictions which do not map back to a valid entity, and then rank the predictions by their log probabilities. Filtering was performed subsequently.
+To evaluate the model, we sample 300 times from the decoder for each input (s,r) pair. We then remove predictions which do not map back to a valid entity, and then rank the predictions by their log probabilities. Filtering was performed subsequently. We achieve 0.22 validation MRR (the full leaderboard is here https://ogb.stanford.edu/docs/lsc/leaderboards/#wikikg90mv2)
 
 You can try the following code in an ipython notebook to evaluate the pre-trained model. The full procedure of mapping entity to ids, filtering etc. is not included here for sake of simplicity but can be provided on request if needed. Please contact Apoorv (apoorvumang@gmail.com) for clarifications/details.
 ---------