Fixed based on review recoemmendation

megatron/arguments.py

@@ -636,10 +636,6 @@ def _add_data_args(parser):
                        '1) a single data path, 2) multiple datasets in the'
                        'form: dataset1-weight dataset1-path dataset2-weight '
                        'dataset2-path ...')
-    group.add_argument('--qa-data-dev', type=str, default=None,
-                       help='Path to the QA dataset dev file.')
-    group.add_argument('--qa-data-test', type=str, default=None,
-                       help='Path to the QA dataset test file.')
     group.add_argument('--split', type=str, default='969, 30, 1',
                        help='Comma-separated list of proportions for training,'
                        ' validation, and test split. For example the split '
@@ -743,18 +739,11 @@ def _add_biencoder_args(parser):
                         'square root of hidden size')
 
     # faiss index
-    group.add_argument('--faiss-use-gpu', action='store_true',
-                       help='Whether create the FaissMIPSIndex on GPU')
     group.add_argument('--block-data-path', type=str, default=None,
                        help='Where to save/load BlockData to/from')
     group.add_argument('--embedding-path', type=str, default=None,
                        help='Where to save/load Open-Retrieval Embedding'
                         ' data to/from')
-    group.add_argument('--faiss-match', type=str, default='string', \
-                        choices=['regex', 'string'], help="Answer matching '\
-                        'logic type")
-    group.add_argument('--faiss-topk-retrievals', type=int, default=100,
-                       help='Number of blocks to use as top-k during retrieval')
 
     # indexer
     group.add_argument('--indexer-batch-size', type=int, default=128,

tasks/main.py

@@ -47,6 +47,20 @@ def get_tasks_args(parser):
                        help='Sliding window for overlapping evaluation.')
     group.add_argument('--strict-lambada', action='store_true',
                        help='Use more difficult formulation of lambada.')
+    # Retriever args
+    group.add_argument('--qa-data-dev', type=str, default=None,
+                       help='Path to the QA dataset dev file.')
+    group.add_argument('--qa-data-test', type=str, default=None,
+                       help='Path to the QA dataset test file.')
+
+    # Faiss arguments for retriever
+    group.add_argument('--faiss-use-gpu', action='store_true',
+                       help='Whether create the FaissMIPSIndex on GPU')
+    group.add_argument('--faiss-match', type=str, default='string', \
+                        choices=['regex', 'string'], help="Answer matching '\
+                        'logic type")
+    group.add_argument('--faiss-topk-retrievals', type=int, default=100,
+                       help='Number of blocks to use as top-k during retrieval')
 
     return parser
 

tasks/orqa/evaluate_orqa.py

@@ -19,8 +19,6 @@ import os
 import sys
 
 from megatron import get_args
-from megatron.initialize import initialize_megatron
-
 from tasks.orqa.evaluate_utils import ORQAEvaluator
 
 def main():

tasks/orqa/natural_questions/qa_utils.py

@@ -2,8 +2,11 @@
 # Copyright (c) Facebook, Inc. and its affiliates.
 # All rights reserved.
 #
-# This source code is licensed under the license found in the
-# LICENSE file in the root directory of this source tree.
+
+# The following code has been taken from
+# https://github.com/facebookresearch/DPR, which is CC-BY-NC 4.0
+# licensed as of now. More details on the license can be found
+# at https://github.com/facebookresearch/DPR/blob/master/LICENSE
 
 """
  Set of utilities for Q&A results validation tasks - Retriver passage

tasks/orqa/natural_questions/tokenizers.py

@@ -2,9 +2,11 @@
 # Copyright (c) Facebook, Inc. and its affiliates.
 # All rights reserved.
 #
-# This source code is licensed under the license found in the
-# LICENSE file in the root directory of this source tree.
 
+# The following code has been taken from
+# https://github.com/facebookresearch/DPR, which is CC-BY-NC 4.0
+# licensed as of now. More details on the license can be found
+# at https://github.com/facebookresearch/DPR/blob/master/LICENSE
 
 """
 Most of the tokenizers code here is copied from DrQA codebase to avoid adding extra dependency