Inherent Challenges of Post-Hoc Membership Inference for Large Language Models
Paper • 2406.17975 • Published
text stringlengths 28.7k 745k | label int64 0 1 |
|---|---|
---
abstract: 'The minimality of the penalty function associated with a convex risk measure is analyzed in this paper. First, in a general static framework, we provide necessary and sufficient conditions for a penalty function defined in a convex and closed subset of the absolutely continuous measures with respect to s... | 1 |
**New Penrose Limits and AdS/CFT**
1.8cm
0.5cm
*$^1$ Dipartimento di Fisica, Università di Perugia,\
I.N.F.N. Sezione di Perugia,\
Via Pascoli, I-06123 Perugia, Italy\
0.4cm *$^2$ NORDITA\
Roslagstullsbacken 23, SE-106 91 Stockholm, Sweden 0.4cm *$^3$ The Niels Bohr Institute\
*Blegdamsvej 17, DK-2100 Copenhagen ... | 1 |
"---\nabstract: 'In this paper, we consider femtocell CR networks, where femto base stations (FBS) a(...TRUNCATED) | 1 |
"---\nabstract: 'We have analyzed available optical data for Au in the mid-infrared range which is i(...TRUNCATED) | 1 |
"---\nabstract: 'We study strong-coupling lattice QCD with staggered-Wilson fermions, with emphasis (...TRUNCATED) | 1 |
"---\nabstract: |\n A non-uniform hypergraph $H=(V,E)$ consists of a vertex set $V$ and an edge s(...TRUNCATED) | 1 |
"---\nauthor:\n- 'Mark Bun [^1]'\n- 'Roi Livni [^2]'\n- 'Shay Moran [^3]'\nbibliography:\n- 'biblio.(...TRUNCATED) | 1 |
"---\nabstract: 'We present high performance implementations of the QR and the singular value decomp(...TRUNCATED) | 1 |
"---\nabstract: 'Multiplicity fluctuations in limited segments of momentum space are calculated for (...TRUNCATED) | 1 |
"---\nabstract: 'The profiles of the chromo-electric field generated by static quark-antiquark, $Q{\(...TRUNCATED) | 1 |
End of preview. Expand in Data Studio
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. 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.
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. 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.
- Downloads last month
- 21