Create subsets to separate long answers, short answer and either if available

In the original dataset, not all questions have answers available. A good amount of questions do not have answers in the context.
To train a model without a bias on "Not available" answers, it would be nice to have subsets with only available answers.

Hence we create 3 subsets:
1. long: Will only have long answer
2. short: Will only have short answer
3. either: Will have one of long/short answers based on availability. In case both are present, long answer will be preferred.

filternq.py

@@ -0,0 +1,143 @@
+"""
+long:
+
+{
+    "document": "",
+    "question": "",
+    "long_answer_candidates": ["", "", ""],
+    "long_answer_candidate_index": 0
+}
+
+short: 
+
+{
+    "document": "",
+    "question": "",
+    "short_answer": ""
+}
+
+either:
+
+{
+    "document": "",
+    "question": "",
+    "answer": ""
+}
+
+"""
+
+
+import sys
+import jsonlines
+from datasets import load_dataset
+from huggingface_hub import HfApi
+
+
+def filter(raw, short_path, long_path, either_path):
+    fps = open(short_path, "a")
+    writers = jsonlines.Writer(fps)
+
+    fpl = open(long_path, "a")
+    writerl = jsonlines.Writer(fpl)
+
+    fpe = open(either_path, "a")
+    writere = jsonlines.Writer(fpe)
+
+    count = 0
+    long = []
+    short = []
+    either = []
+    for sample in raw:
+        try:
+            answer = ""
+
+            if sample["short_answers"][0]:
+                answer = sample["short_answers"][0]
+                
+                short.append({
+                    "document": sample["document"],
+                    "question": sample["question"],
+                    "short_answer": answer
+                })
+
+            if sample["long_answer_candidate_index"] != -1:
+                answer = sample["long_answer_candidates"][sample["long_answer_candidate_index"]]        # long answer will have precedence over short answer
+                
+                long.append({
+                    "document": sample["document"],
+                    "question": sample["question"],
+                    "long_answer_candidates": sample["long_answer_candidates"],
+                    "long_answer_candidate_index": sample["long_answer_candidate_index"]
+                })
+
+            if answer:
+                count += 1          # count only if there is an answer
+                
+                either.append({
+                    "document": sample["document"],
+                    "question": sample["question"],
+                    "answer": answer
+                })
+            
+        except Exception as ex:
+            # raise ex
+            print("Exception: " + str(ex))
+
+        if (count + 1) % 1000 == 0:
+            writere.write_all(either)
+            either = []
+
+            if short:
+                writers.write_all(short)
+                short = []
+            
+            if long:
+                writerl.write_all(long)
+                long = []
+
+        print("Done: " + str(count), end="\r")
+
+
+    if either:
+        writere.write_all(either)
+        either = []
+
+    if short:
+        writers.write_all(short)
+        short = []
+    
+    if long:
+        writerl.write_all(long)
+        long = []
+
+    writere.close()
+    fpe.close()
+
+    writers.close()
+    fps.close()
+
+    writerl.close()
+    fpl.close()
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 1:
+        raise AttributeError("Missing required argument: repository id")
+
+    repo = sys.argv[1]
+
+    api = HfApi()
+
+    train_data = load_dataset(repo, split="train", streaming=True)
+    filter(raw=train_data, short_path="data/short/train.jsonl", long_path="data/long/train.jsonl", either_path="data/either/train.jsonl")
+
+    val_data = load_dataset(repo, split="validation", streaming=True)
+    filter(raw=val_data, short_path="data/short/validation.jsonl", long_path="data/long/validation.jsonl", either_path="data/either/validation.jsonl")
+
+    api.upload_folder(
+        folder_path="data/",
+        repo_id=repo,
+        repo_type="dataset",
+        multi_commits=True,
+        multi_commits_verbose=True
+    )