Datasets:

ArXiv:
Maurice Weber commited on
Commit
8b3730f
1 Parent(s): 167a6e2

add minhash to table

Browse files
Files changed (1) hide show
  1. README.md +17 -13
README.md CHANGED
@@ -78,7 +78,8 @@ A full set of scripts to recreate the dataset, including the quality signals, ca
78
  found [here](https://github.com/togethercomputer/RedPajama-Data).
79
 
80
  ### Applying Filtering Rules
81
- You can use the quality signals to filter the raw RedPajama-V2 dataset for a given set of rules. For example, consider
 
82
  the following set of rules used in Gopher:
83
 
84
  ```python
@@ -98,7 +99,7 @@ def gopher_rules_pass(sample) -> bool:
98
 
99
  # rule 2: symbol to word ratio below 0.1
100
  symbol_word_ratio = signals["rps_doc_symbol_to_word_ratio"][0][2]
101
- if symbol_word_ratio > 0.1:
102
  return False
103
 
104
  # rule 3: 90% of lines need to start without a bullet point
@@ -114,8 +115,7 @@ def gopher_rules_pass(sample) -> bool:
114
  return False
115
 
116
  # rule 5: ...
117
-
118
-
119
  return True
120
  ```
121
 
@@ -123,21 +123,21 @@ Filtering the RedPajama-V2 dataset with this set of rules is then as easy as:
123
 
124
  ```python
125
  ds_iterator = load_dataset(
126
- "togethercomputer/RedPajama-Data-V2",
127
- snapshots=["2023-14"],
128
- languages=["en"],
129
- name="default",
130
  streaming=True
131
  )
132
 
133
  filtered_dataset = []
134
 
135
  for sample in ds_iterator["train"]:
136
-
137
- if not gopher_rules_pass(sample):
138
- continue
139
-
140
- filtered_dataset.append(sample)
141
  ```
142
 
143
  ### Dataset Summary
@@ -191,6 +191,10 @@ RedPajama-V2 is an open dataset for training large laguage models and includes o
191
  | rps_doc_frac_chars_top_4gram | The fraction of characters in the top word 4gram. | Repetitiveness | [RefinedWeb](https://arxiv.org/abs/2306.01116), [Gopher](https://arxiv.org/abs/2112.11446) |
192
  | rps_doc_ldnoobw_words | The number of sequences of words that are contained in the List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words blocklist. The blocklist is obtained from the [LDNOOBW](https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words) repo. | toxicity | [C4](https://arxiv.org/abs/1910.10683) |
193
  | rps_doc_ut1_blacklist | A categorical id corresponding to the list of categories of the domain of the document. Categories are obtained from the UT1 blacklist. The list is obtained from [UT-Capitole](https://dsi.ut-capitole.fr/blacklists/). | toxicictiy | [RefinedWeb](https://arxiv.org/abs/2306.01116) |
 
 
 
 
194
 
195
  #### Document and Token Counts for the Annotated and deduplicated `head_middle` part of the dataset
196
 
 
78
  found [here](https://github.com/togethercomputer/RedPajama-Data).
79
 
80
  ### Applying Filtering Rules
81
+
82
+ You can use the quality signals to filter the raw RedPajama-V2 dataset for a given set of rules. For example, consider
83
  the following set of rules used in Gopher:
84
 
85
  ```python
 
99
 
100
  # rule 2: symbol to word ratio below 0.1
101
  symbol_word_ratio = signals["rps_doc_symbol_to_word_ratio"][0][2]
102
+ if symbol_word_ratio > 0.1:
103
  return False
104
 
105
  # rule 3: 90% of lines need to start without a bullet point
 
115
  return False
116
 
117
  # rule 5: ...
118
+
 
119
  return True
120
  ```
121
 
 
123
 
124
  ```python
125
  ds_iterator = load_dataset(
126
+ "togethercomputer/RedPajama-Data-V2",
127
+ snapshots=["2023-14"],
128
+ languages=["en"],
129
+ name="default",
130
  streaming=True
131
  )
132
 
133
  filtered_dataset = []
134
 
135
  for sample in ds_iterator["train"]:
136
+
137
+ if not gopher_rules_pass(sample):
138
+ continue
139
+
140
+ filtered_dataset.append(sample)
141
  ```
142
 
143
  ### Dataset Summary
 
191
  | rps_doc_frac_chars_top_4gram | The fraction of characters in the top word 4gram. | Repetitiveness | [RefinedWeb](https://arxiv.org/abs/2306.01116), [Gopher](https://arxiv.org/abs/2112.11446) |
192
  | rps_doc_ldnoobw_words | The number of sequences of words that are contained in the List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words blocklist. The blocklist is obtained from the [LDNOOBW](https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words) repo. | toxicity | [C4](https://arxiv.org/abs/1910.10683) |
193
  | rps_doc_ut1_blacklist | A categorical id corresponding to the list of categories of the domain of the document. Categories are obtained from the UT1 blacklist. The list is obtained from [UT-Capitole](https://dsi.ut-capitole.fr/blacklists/). | toxicictiy | [RefinedWeb](https://arxiv.org/abs/2306.01116) |
194
+ | minhash_signature_0.7 | Banded minhash signature of the document, for fuzzy deduplication at Jaccard similarity 0.7 | Deduplication |
195
+ | minhash_signature_0.8 | Banded minhash signature of the document, for fuzzy deduplication at Jaccard similarity 0.8 | Deduplication |
196
+ | minhash_signature_0.9 | Banded minhash signature of the document, for fuzzy deduplication at Jaccard similarity 0.9 | Deduplication |
197
+ | minhash_signature_1.0 | Banded minhash signature of the document, for fuzzy deduplication at Jaccard similarity 1.0 | Deduplication |
198
 
199
  #### Document and Token Counts for the Annotated and deduplicated `head_middle` part of the dataset
200