Poisoning Web-Scale Training Datasets is Practical · HackerPapers discussions

Poisoning Web-Scale Training Datasets is Practical — mentioned by Terr_

Fri, 21 Mar 2025 00:41:35 GMT

> LLMs are already trained on a lot of dubious sources (for example Quora and Reddit) There's a very big difference between crawling documents where people happen to say something wrong (e.g. "Tortoises are amphibians") versus the impact of widespread poisoning [0] or indirect prompt injection. [1] [0] https://arxiv.org/html/2302.10149v2 [1] https://cetas.turing.ac.uk/publications/indirect-prompt-inje...

Paper and discussion history. A mention is not an endorsement.

Poisoning Web-Scale Training Datasets is Practical — mentioned by Blahah

Sun, 04 Jun 2023 22:20:30 GMT

https://arxiv.org/abs/2302.10149 maybe

Paper and discussion history. A mention is not an endorsement.

Poisoning Web-Scale Training Datasets is Practical — mentioned by PaulHoule

Wed, 05 Apr 2023 11:13:41 GMT

see https://arxiv.org/abs/2302.10149

Paper and discussion history. A mention is not an endorsement.

Poisoning Web-Scale Training Datasets is Practical — HN submission

Tue, 21 Feb 2023 23:36:41 GMT

Poisoning web-scale training datasets is practical

171 points; 94 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Poisoning Web-Scale Training Datasets is Practical — HN submission

Tue, 21 Feb 2023 12:00:47 GMT

Poisoning Web-Scale Training Datasets Is Practical

4 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.