Has this LLM been pretrained on this corpus?

[lintool]

Let's try to characterize data pollution - i.e., has this LLM been pretrained on this corpus?

Simple task: pick a random passage - chop into half. Feed first half into LLM, ask it to complete the passage.

[lintool]

Relevant paper: https://arxiv.org/abs/2311.01964

[AndreSlavescu]

Training Data Extraction:

https://arxiv.org/abs/2311.17035

[ASChampOmega]

Early Literature Review:

Literature Review for Data Contamination.docx

[lintool]

This is the first passage in the MS MARCO passage corpus:

{"id": "0", "contents": "The presence of communication amid scientific minds was equally important to the success of the Manhattan Project as scientific intellect was. The only cloud hanging over the impressive achievement of the atomic researchers and engineers is what their success truly meant; hundreds of thousands of innocent lives obliterated."}

Here's me playing around... https://chat.openai.com/share/8b19eeec-4be2-4a16-a1ad-12b68fad81f2

[lintool]

Investigating Data Contamination in Modern Benchmarks for Large Language Models
https://arxiv.org/abs/2311.09783

[crystina-z]

some related papers:

• Membership Inference Attacks against Machine Learning Models (2017)
• LOGAN: Membership Inference Attacks Against Generative Models (2017)
• Auditing Data Provenance in Text-Generation Models (2019)
• Demystifying Membership Inference Attacks in Machine Learning as a Service (2019)
• Membership Inference Attacks on Machine Learning: A Survey (2022)