updated links

README.md

@@ -24,7 +24,7 @@ We use a combination of three French Q&A datasets:
 
 We are using 90 562 random questions for `train` and 22 391 for `dev`. No question in `train` exists in `dev`. For each question, we have a single `positive_context` (the paragraph where the answer to this question is found) and around 30 `hard_negtive_contexts`. Hard negative contexts are found by querying an ES instance (via bm25 retrieval) and getting the top-k candidates **that do not contain the answer**. 
 
-The corresponding files are here: 
+The files are over [here](https://drive.google.com/file/d/1W5Jm3sqqWlsWsx2sFpA39Ewn33PaLQ7U/view?usp=sharing). 
 
 ### Evaluation
 
@@ -33,7 +33,7 @@ We use FQuADv1.0 and French-SQuAD evaluation sets.
 
 
 ## Training Script
-We use the official [Facebook DPR implentation](https://github.com/facebookresearch/DPR) with a slight modification: by default, the code can work with Roberta models, still we changed a single line to make it easier to work with Camembert. This modification can be found [over here](dpr fork).
+We use the official [Facebook DPR implentation](https://github.com/facebookresearch/DPR) with a slight modification: by default, the code can work with Roberta models, still we changed a single line to make it easier to work with Camembert. This modification can be found [over here](https://github.com/psorianom/DPR).
 
 ### Hyperparameters