Datasets:

Modalities:
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
pandas
File size: 1,529 Bytes
0bd4f69
 
 
 
 
 
 
 
 
 
 
 
 
 
b9a69e5
fa54622
6484ff0
3246a85
3561451
54ea011
ecddadc
 
 
 
3246a85
243613e
3246a85
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: label
    dtype: int64
  splits:
  - name: train
    num_bytes: 146613669
    num_examples: 2000
  download_size: 67134534
  dataset_size: 146613669
---
# ArXiv papers from The Pile for document-level MIAs against LLMs

This dataset contains **full** ArXiv papers randomly sampled from the train (members) and test (non-members) dataset from (the uncopyrighted version of) [the Pile](https://huggingface.co/datasets/monology/pile-uncopyrighted). 
We randomly sample 1,000 documents members and 1,000 non-members, ensuring that the selected documents have at least 5,000 words (any sequences of characters seperated by a white space). 
We also provide the dataset where each document is split into 25 sequences of 200 words [here](https://huggingface.co/datasets/imperial-cpg/pile_arxiv_doc_mia_sequences).

The dataset contains as columns: 
- text: the raw text of the sequence
- label: binary label for membership (1=member)

The dataset can be used to develop and evaluate document-level MIAs against LLMs trained on The Pile. 
Target models include the suite of Pythia and GPTNeo models, to be found [here](https://huggingface.co/EleutherAI). Our understanding is that the deduplication executed on the Pile to create the "Pythia-dedup" models has been only done on the training dataset, suggesting this dataset of members/non-members also to be valid for these models. 

For more information we refer to [the paper](https://arxiv.org/pdf/2406.17975).