Datasets:
Maurice Weber
commited on
Commit
•
2cae54f
1
Parent(s):
04e7975
add dedupe flag
Browse files- RedPajama-Data-V2.py +61 -13
RedPajama-Data-V2.py
CHANGED
@@ -23,6 +23,8 @@ import os
|
|
23 |
import gzip
|
24 |
from typing import List
|
25 |
|
|
|
|
|
26 |
logger = datasets.logging.get_logger(__name__)
|
27 |
|
28 |
_DESCRIPTION = """\
|
@@ -188,13 +190,22 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
188 |
]
|
189 |
})
|
190 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
return [
|
192 |
datasets.SplitGenerator(
|
193 |
name=datasets.Split.TRAIN,
|
194 |
gen_kwargs={
|
195 |
"listings_ids": {"head_middle": listings},
|
196 |
"documents_files": documents_files,
|
197 |
-
"quality_signals_files": quality_signals_files
|
|
|
198 |
}
|
199 |
)
|
200 |
]
|
@@ -235,16 +246,21 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
235 |
line.strip() for line in f
|
236 |
])
|
237 |
|
238 |
-
# build urls pointing to documents and
|
239 |
document_urls = {}
|
240 |
quality_signals_urls = {}
|
|
|
241 |
for part, part_listings_ids in listings_ids.items():
|
|
|
242 |
document_urls[part] = []
|
243 |
quality_signals_urls[part] = []
|
|
|
|
|
244 |
for lst_id in part_listings_ids:
|
245 |
document_urls[part].append(
|
246 |
os.path.join(_URL_BASE, f"documents/{lst_id}.json.gz")
|
247 |
)
|
|
|
248 |
if part != "head_middle":
|
249 |
continue
|
250 |
|
@@ -254,8 +270,15 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
254 |
)
|
255 |
)
|
256 |
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
documents_files = dl_manager.download(document_urls)
|
258 |
quality_signals_files = dl_manager.download(quality_signals_urls)
|
|
|
259 |
|
260 |
return [
|
261 |
datasets.SplitGenerator(
|
@@ -263,7 +286,8 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
263 |
gen_kwargs={
|
264 |
"listings_ids": listings_ids,
|
265 |
"documents_files": documents_files,
|
266 |
-
"quality_signals_files": quality_signals_files
|
|
|
267 |
}
|
268 |
)
|
269 |
]
|
@@ -275,13 +299,15 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
275 |
return self._split_generators_full(dl_manager)
|
276 |
|
277 |
def _generate_examples(
|
278 |
-
self, listings_ids, documents_files, quality_signals_files
|
|
|
279 |
):
|
280 |
key = 0
|
281 |
for part in documents_files.keys():
|
282 |
part_docs_files = documents_files[part]
|
283 |
part_qs_files = quality_signals_files[part]
|
284 |
part_listings_ids = listings_ids[part]
|
|
|
285 |
|
286 |
if len(part_qs_files) == 0:
|
287 |
for sample in self._handle_tail_partition(
|
@@ -292,7 +318,9 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
292 |
continue
|
293 |
|
294 |
for sample in self._handle_head_middle_partition(
|
295 |
-
part, part_docs_files, part_qs_files,
|
|
|
|
|
296 |
):
|
297 |
yield key, sample
|
298 |
key += 1
|
@@ -303,7 +331,7 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
303 |
for row, doc in enumerate(df):
|
304 |
doc_id = f"{listing_id}.json.gz/{row}"
|
305 |
try:
|
306 |
-
yield self.handle_record(part, doc_id, doc, None)
|
307 |
except Exception as e:
|
308 |
print(f'doc_file: {doc_file}')
|
309 |
print(f'row: {row}')
|
@@ -311,22 +339,41 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
311 |
raise e
|
312 |
|
313 |
def _handle_head_middle_partition(
|
314 |
-
self, part, docs_files, qs_files, listings_ids
|
315 |
):
|
316 |
assert len(docs_files) == len(qs_files)
|
317 |
|
318 |
listings_ids = listings_ids[:len(docs_files)]
|
319 |
|
320 |
-
for doc_file, qs_file, listings_id in zip(
|
321 |
-
docs_files, qs_files, listings_ids
|
322 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
323 |
try:
|
324 |
with gzip.open(doc_file, "rt", encoding="utf-8") as df:
|
325 |
with gzip.open(qs_file, "rt", encoding="utf-8") as qf:
|
326 |
for row, (doc, qs) in enumerate(zip(df, qf)):
|
327 |
doc_id = f"{listings_id}.json.gz/{row}"
|
|
|
|
|
|
|
|
|
|
|
|
|
328 |
try:
|
329 |
-
yield self.handle_record(
|
|
|
|
|
|
|
330 |
except Exception as e:
|
331 |
print(f'failed handling row {row} in '
|
332 |
f'{doc_file} ({qs_file})')
|
@@ -339,7 +386,7 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
339 |
continue
|
340 |
|
341 |
@staticmethod
|
342 |
-
def handle_record(part, doc_id, doc, qs):
|
343 |
doc = json.loads(doc)
|
344 |
qs = json.loads(qs) if qs is not None else {}
|
345 |
|
@@ -352,11 +399,12 @@ class RedPajamaV2(datasets.GeneratorBasedBuilder):
|
|
352 |
"digest": doc["digest"],
|
353 |
}
|
354 |
|
355 |
-
quality_signals =
|
|
|
356 |
|
357 |
return {
|
358 |
"raw_content": doc["raw_content"],
|
359 |
"doc_id": doc_id,
|
360 |
"meta": json.dumps(meta),
|
361 |
-
"quality_signals": quality_signals,
|
362 |
}
|
|
|
23 |
import gzip
|
24 |
from typing import List
|
25 |
|
26 |
+
import pyarrow.parquet as pq
|
27 |
+
|
28 |
logger = datasets.logging.get_logger(__name__)
|
29 |
|
30 |
_DESCRIPTION = """\
|
|
|
190 |
]
|
191 |
})
|
192 |
|
193 |
+
# fetch ids of duplicates
|
194 |
+
duplicates_ids_files = dl_manager.download({
|
195 |
+
"head_middle": [
|
196 |
+
f"sample/duplicates/{lst}.duplicates.parquet"
|
197 |
+
for lst in listings
|
198 |
+
]
|
199 |
+
})
|
200 |
+
|
201 |
return [
|
202 |
datasets.SplitGenerator(
|
203 |
name=datasets.Split.TRAIN,
|
204 |
gen_kwargs={
|
205 |
"listings_ids": {"head_middle": listings},
|
206 |
"documents_files": documents_files,
|
207 |
+
"quality_signals_files": quality_signals_files,
|
208 |
+
"duplicates_ids_files": duplicates_ids_files
|
209 |
}
|
210 |
)
|
211 |
]
|
|
|
246 |
line.strip() for line in f
|
247 |
])
|
248 |
|
249 |
+
# build urls pointing to documents, quality signals and duplicate ids
|
250 |
document_urls = {}
|
251 |
quality_signals_urls = {}
|
252 |
+
duplicates_ids_urls = {}
|
253 |
for part, part_listings_ids in listings_ids.items():
|
254 |
+
|
255 |
document_urls[part] = []
|
256 |
quality_signals_urls[part] = []
|
257 |
+
duplicates_ids_urls[part] = []
|
258 |
+
|
259 |
for lst_id in part_listings_ids:
|
260 |
document_urls[part].append(
|
261 |
os.path.join(_URL_BASE, f"documents/{lst_id}.json.gz")
|
262 |
)
|
263 |
+
|
264 |
if part != "head_middle":
|
265 |
continue
|
266 |
|
|
|
270 |
)
|
271 |
)
|
272 |
|
273 |
+
duplicates_ids_urls[part].append(
|
274 |
+
os.path.join(
|
275 |
+
_URL_BASE, f"duplicates/{lst_id}.duplicates.parquet"
|
276 |
+
)
|
277 |
+
)
|
278 |
+
|
279 |
documents_files = dl_manager.download(document_urls)
|
280 |
quality_signals_files = dl_manager.download(quality_signals_urls)
|
281 |
+
duplicates_ids_files = dl_manager.download(duplicates_ids_urls)
|
282 |
|
283 |
return [
|
284 |
datasets.SplitGenerator(
|
|
|
286 |
gen_kwargs={
|
287 |
"listings_ids": listings_ids,
|
288 |
"documents_files": documents_files,
|
289 |
+
"quality_signals_files": quality_signals_files,
|
290 |
+
"duplicates_ids_files": duplicates_ids_files
|
291 |
}
|
292 |
)
|
293 |
]
|
|
|
299 |
return self._split_generators_full(dl_manager)
|
300 |
|
301 |
def _generate_examples(
|
302 |
+
self, listings_ids, documents_files, quality_signals_files,
|
303 |
+
duplicates_ids_files
|
304 |
):
|
305 |
key = 0
|
306 |
for part in documents_files.keys():
|
307 |
part_docs_files = documents_files[part]
|
308 |
part_qs_files = quality_signals_files[part]
|
309 |
part_listings_ids = listings_ids[part]
|
310 |
+
part_duplicates_ids_files = duplicates_ids_files[part]
|
311 |
|
312 |
if len(part_qs_files) == 0:
|
313 |
for sample in self._handle_tail_partition(
|
|
|
318 |
continue
|
319 |
|
320 |
for sample in self._handle_head_middle_partition(
|
321 |
+
part, part_docs_files, part_qs_files,
|
322 |
+
part_duplicates_ids_files, part_listings_ids
|
323 |
+
|
324 |
):
|
325 |
yield key, sample
|
326 |
key += 1
|
|
|
331 |
for row, doc in enumerate(df):
|
332 |
doc_id = f"{listing_id}.json.gz/{row}"
|
333 |
try:
|
334 |
+
yield self.handle_record(part, doc_id, doc, None, None)
|
335 |
except Exception as e:
|
336 |
print(f'doc_file: {doc_file}')
|
337 |
print(f'row: {row}')
|
|
|
339 |
raise e
|
340 |
|
341 |
def _handle_head_middle_partition(
|
342 |
+
self, part, docs_files, qs_files, dupes_files, listings_ids,
|
343 |
):
|
344 |
assert len(docs_files) == len(qs_files)
|
345 |
|
346 |
listings_ids = listings_ids[:len(docs_files)]
|
347 |
|
348 |
+
for doc_file, qs_file, dupe_file, listings_id in zip(
|
349 |
+
docs_files, qs_files, dupes_files, listings_ids
|
350 |
):
|
351 |
+
# load duplicates
|
352 |
+
try:
|
353 |
+
with open(dupe_file, "rb") as df:
|
354 |
+
duplicates = set(pq.read_table(
|
355 |
+
df, columns=["doc_id"], use_pandas_metadata=False
|
356 |
+
)["doc_id"].to_pylist())
|
357 |
+
except Exception as e:
|
358 |
+
print(f'failed loading duplicate ids from {dupe_file}.')
|
359 |
+
duplicates = set()
|
360 |
+
|
361 |
try:
|
362 |
with gzip.open(doc_file, "rt", encoding="utf-8") as df:
|
363 |
with gzip.open(qs_file, "rt", encoding="utf-8") as qf:
|
364 |
for row, (doc, qs) in enumerate(zip(df, qf)):
|
365 |
doc_id = f"{listings_id}.json.gz/{row}"
|
366 |
+
|
367 |
+
if doc_id in duplicates:
|
368 |
+
is_duplicate = True
|
369 |
+
else:
|
370 |
+
is_duplicate = False
|
371 |
+
|
372 |
try:
|
373 |
+
yield self.handle_record(
|
374 |
+
part, doc_id, doc, qs,
|
375 |
+
is_duplicate=is_duplicate
|
376 |
+
)
|
377 |
except Exception as e:
|
378 |
print(f'failed handling row {row} in '
|
379 |
f'{doc_file} ({qs_file})')
|
|
|
386 |
continue
|
387 |
|
388 |
@staticmethod
|
389 |
+
def handle_record(part, doc_id, doc, qs, is_duplicate=None):
|
390 |
doc = json.loads(doc)
|
391 |
qs = json.loads(qs) if qs is not None else {}
|
392 |
|
|
|
399 |
"digest": doc["digest"],
|
400 |
}
|
401 |
|
402 |
+
quality_signals = qs.get("quality_signals", {})
|
403 |
+
quality_signals["is_duplicate"] = is_duplicate
|
404 |
|
405 |
return {
|
406 |
"raw_content": doc["raw_content"],
|
407 |
"doc_id": doc_id,
|
408 |
"meta": json.dumps(meta),
|
409 |
+
"quality_signals": json.dumps(quality_signals),
|
410 |
}
|