Hugging Face's logo

Datasets:
RMT-team
/
babilong-train-5k-samples

Modalities:
Text
Size:
100K - 1M
ArXiv:
Dataset card Viewer Files Community
2
booydar commited on
Commit
79dfe51
·
verified ·
1 Parent(s): 8083a82

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +205 -0
README.md ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ configs:
3
+ - config_name: 0k
4
+ data_files:
5
+ - split: qa1
6
+ path: data/qa1/0k.json
7
+ - split: qa2
8
+ path: data/qa2/0k.json
9
+ - split: qa3
10
+ path: data/qa3/0k.json
11
+ - split: qa4
12
+ path: data/qa4/0k.json
13
+ - split: qa5
14
+ path: data/qa5/0k.json
15
+ - split: qa6
16
+ path: data/qa6/0k.json
17
+ - split: qa7
18
+ path: data/qa7/0k.json
19
+ - split: qa8
20
+ path: data/qa8/0k.json
21
+ - split: qa9
22
+ path: data/qa9/0k.json
23
+ - split: qa10
24
+ path: data/qa10/0k.json
25
+ - config_name: 1k
26
+ data_files:
27
+ - split: qa1
28
+ path: data/qa1/1k.json
29
+ - split: qa2
30
+ path: data/qa2/1k.json
31
+ - split: qa3
32
+ path: data/qa3/1k.json
33
+ - split: qa4
34
+ path: data/qa4/1k.json
35
+ - split: qa5
36
+ path: data/qa5/1k.json
37
+ - split: qa6
38
+ path: data/qa6/1k.json
39
+ - split: qa7
40
+ path: data/qa7/1k.json
41
+ - split: qa8
42
+ path: data/qa8/1k.json
43
+ - split: qa9
44
+ path: data/qa9/1k.json
45
+ - split: qa10
46
+ path: data/qa10/1k.json
47
+ - config_name: 2k
48
+ data_files:
49
+ - split: qa1
50
+ path: data/qa1/2k.json
51
+ - split: qa2
52
+ path: data/qa2/2k.json
53
+ - split: qa3
54
+ path: data/qa3/2k.json
55
+ - split: qa4
56
+ path: data/qa4/2k.json
57
+ - split: qa5
58
+ path: data/qa5/2k.json
59
+ - split: qa6
60
+ path: data/qa6/2k.json
61
+ - split: qa7
62
+ path: data/qa7/2k.json
63
+ - split: qa8
64
+ path: data/qa8/2k.json
65
+ - split: qa9
66
+ path: data/qa9/2k.json
67
+ - split: qa10
68
+ path: data/qa10/2k.json
69
+ - config_name: 4k
70
+ data_files:
71
+ - split: qa1
72
+ path: data/qa1/4k.json
73
+ - split: qa2
74
+ path: data/qa2/4k.json
75
+ - split: qa3
76
+ path: data/qa3/4k.json
77
+ - split: qa4
78
+ path: data/qa4/4k.json
79
+ - split: qa5
80
+ path: data/qa5/4k.json
81
+ - split: qa6
82
+ path: data/qa6/4k.json
83
+ - split: qa7
84
+ path: data/qa7/4k.json
85
+ - split: qa8
86
+ path: data/qa8/4k.json
87
+ - split: qa9
88
+ path: data/qa9/4k.json
89
+ - split: qa10
90
+ path: data/qa10/4k.json
91
+ - config_name: 8k
92
+ data_files:
93
+ - split: qa1
94
+ path: data/qa1/8k.json
95
+ - split: qa2
96
+ path: data/qa2/8k.json
97
+ - split: qa3
98
+ path: data/qa3/8k.json
99
+ - split: qa4
100
+ path: data/qa4/8k.json
101
+ - split: qa5
102
+ path: data/qa5/8k.json
103
+ - split: qa6
104
+ path: data/qa6/8k.json
105
+ - split: qa7
106
+ path: data/qa7/8k.json
107
+ - split: qa8
108
+ path: data/qa8/8k.json
109
+ - split: qa9
110
+ path: data/qa9/8k.json
111
+ - split: qa10
112
+ path: data/qa10/8k.json
113
+ - config_name: 16k
114
+ data_files:
115
+ - split: qa1
116
+ path: data/qa1/16k.json
117
+ - split: qa2
118
+ path: data/qa2/16k.json
119
+ - split: qa3
120
+ path: data/qa3/16k.json
121
+ - split: qa4
122
+ path: data/qa4/16k.json
123
+ - split: qa5
124
+ path: data/qa5/16k.json
125
+ - split: qa6
126
+ path: data/qa6/16k.json
127
+ - split: qa7
128
+ path: data/qa7/16k.json
129
+ - split: qa8
130
+ path: data/qa8/16k.json
131
+ - split: qa9
132
+ path: data/qa9/16k.json
133
+ - split: qa10
134
+ path: data/qa10/16k.json
135
+ - config_name: 32k
136
+ data_files:
137
+ - split: qa1
138
+ path: data/qa1/32k.json
139
+ - split: qa2
140
+ path: data/qa2/32k.json
141
+ - split: qa3
142
+ path: data/qa3/32k.json
143
+ - split: qa4
144
+ path: data/qa4/32k.json
145
+ - split: qa5
146
+ path: data/qa5/32k.json
147
+ - split: qa6
148
+ path: data/qa6/32k.json
149
+ - split: qa7
150
+ path: data/qa7/32k.json
151
+ - split: qa8
152
+ path: data/qa8/32k.json
153
+ - split: qa9
154
+ path: data/qa9/32k.json
155
+ - split: qa10
156
+ path: data/qa10/32k.json
157
+ ---
158
+
159
+ # BABILong (5k train samples) : a long-context needle-in-a-haystack benchmark for LLMs
160
+
161
+ Preprint is on [arXiv](https://arxiv.org/abs/2402.10790)
162
+
163
+ ## bAbI + Books = BABILong
164
+
165
+ **BABILong** is a novel generative benchmark for evaluating the performance of NLP models in
166
+ processing arbitrarily long documents with distributed facts.
167
+
168
+ It contains 10 configs, each corresponding to its bAbI task. Each config has spltis corresponding to different sequence lengths in tokens: '4k', '32k', '128k', '256k', '512k', '1M'
169
+
170
+ Solving tasks with a long context size requires the model to distinguish important information from large amounts of irrelevant details. To simulate this behavior we ”hide” the sentences of the original task between the sentences of irrelevant text. We use the [bAbI](https://huggingface.co/datasets/facebook/babi_qa) dataset [1] as facts and [PG19](https://huggingface.co/datasets/pg19) as background text. Resulting test samples might have lenghts of **millions of tokens**.
171
+
172
+ BABILong consists of 10 tasks designed for evaluation of basic aspects of reasoning. The bAbI tasks are generated by simulating a set of characters and objects engaged in various movements and interactions with each other in multiple locations. Each interaction is represented by a fact, e.g. **”Mary travelled to the office”**, and the task is to answer a question using the facts from the current simulation, for instance, **”Where is Mary?”**. The bAbI tasks vary based on the number of facts, question complexity and the aspects of reasoning.
173
+
174
+ ### First ten tasks of BABILong
175
+
176
+ | Task | Name | facts per task | supporting facts per task |
177
+ |------|--------------------------|-----------------|---------------------------|
178
+ | qa1 | single supporting fact | 2 - 10 | 1 |
179
+ | qa2 | two supporting facts | 2 - 68 | 2 |
180
+ | qa3 | three supporting facts | 4 - 32 | 3 |
181
+ | qa4 | two arg relations | 2 | 1 |
182
+ | qa5 | three arg relations | 2 - 126 | 1 |
183
+ | qa6 | yes-no questions | 2 - 26 | 1 |
184
+ | qa7 | counting | 2 - 52 | 1-10 |
185
+ | qa8 | lists-sets | 2 - 50 | 1-8 |
186
+ | qa9 | simple negation | 2 - 10 | 1 |
187
+ | qa10 | indefinite knowledge | 2 - 10 | 1 |
188
+
189
+ Join us in this exciting endeavor and let's push the boundaries of what's possible together!
190
+
191
+ ## Citation
192
+ ```
193
+ @misc{kuratov2024search,
194
+ title={In Search of Needles in a 10M Haystack: Recurrent Memory Finds What LLMs Miss},
195
+ author={Yuri Kuratov and Aydar Bulatov and Petr Anokhin and Dmitry Sorokin and Artyom Sorokin and Mikhail Burtsev},
196
+ year={2024},
197
+ eprint={2402.10790},
198
+ archivePrefix={arXiv},
199
+ primaryClass={cs.CL}
200
+ }
201
+ ```
202
+
203
+ ## References
204
+
205
+ [1] Weston, Jason, et al. "Towards ai-complete question answering: A set of prerequisite toy tasks." arXiv preprint [arXiv:1502.05698](https://arxiv.org/abs/1502.05698) (2015).