Support streaming (#3)

- Use iter_archive for TAR file (5877ce5051fbbf4f7e9a67f8b2445ab6bde9e0fb)

the_pile_openwebtext2.py

@@ -14,11 +14,8 @@
 # limitations under the License.
 """The OpenWebText2 Corpus"""
 
-
-import glob
 import io
 import json
-import os
 
 import zstandard
 
@@ -70,18 +67,22 @@ class Openwebtext2(datasets.GeneratorBasedBuilder):
         )
 
     def _split_generators(self, dl_manager):
-        dl_dir = dl_manager.download_and_extract(_URL)
-        files = glob.glob(os.path.join(dl_dir, "*jsonl.zst"))
+        archive = dl_manager.download(_URL)
         return [
-            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"files": files}),
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={"files": dl_manager.iter_archive(archive)},
+            ),
         ]
 
     def _generate_examples(self, files):
         """Yields examples."""
         _id = 0
-        for file_path in files:
+        for path, file in files:
+            if not path.endswith(".jsonl.zst"):
+                continue
             reader = Reader()
-            for document, metadata in reader.read_jsonl(file_path, get_meta=True):
+            for document, metadata in reader.read_jsonl(file, get_meta=True):
                 yield _id, {
                     "title": metadata["title"],
                     "text": document,
@@ -95,25 +96,24 @@ class Reader:
     def __init__(self):
         pass
 
-    def read_jsonl(self, file, get_meta=False, autojoin_paragraphs=True, para_joiner="\n\n"):
-        with open(file, "rb") as fh:
-            self.fh = fh
-            cctx = zstandard.ZstdDecompressor()
-            reader = io.BufferedReader(cctx.stream_reader(fh))
-            for line in reader:
-                ob = json.loads(line)
-                # naive jsonl where each object is just the string itself, with no meta. For legacy compatibility.
-                if isinstance(ob, str):
-                    assert not get_meta
-                    yield ob
-                    continue
-
-                text = ob["text"]
-
-                if autojoin_paragraphs and isinstance(text, list):
-                    text = para_joiner.join(text)
-
-                if get_meta:
-                    yield text, (ob["meta"] if "meta" in ob else {})
-                else:
-                    yield text
+    def read_jsonl(self, fh, get_meta=False, autojoin_paragraphs=True, para_joiner="\n\n"):
+        self.fh = fh
+        cctx = zstandard.ZstdDecompressor()
+        reader = io.BufferedReader(cctx.stream_reader(fh))
+        for line in reader:
+            ob = json.loads(line)
+            # naive jsonl where each object is just the string itself, with no meta. For legacy compatibility.
+            if isinstance(ob, str):
+                assert not get_meta
+                yield ob
+                continue
+
+            text = ob["text"]
+
+            if autojoin_paragraphs and isinstance(text, list):
+                text = para_joiner.join(text)
+
+            if get_meta:
+                yield text, (ob["meta"] if "meta" in ob else {})
+            else:
+                yield text