Author: Alex Turner. Contributors: Dipika Khullar, Ed Turner, and Roy Rinberg.

Dataset contamination is bad for several reasons. Most obviously, when benchmarks are included in AI training data, those benchmarks no longer measure generalization -- the AI may have been directly taught the answers. Even more concerningly, if your data promote negative "stereotypes" about AIs, they might become self-fulfilling prophecies, training future models to exhibit those very behaviors.

In the Claude 4 system card, Anthropic revealed that approximately 250,000 transcripts from their alignment faking paper had been scraped from the public web and included in their pretraining data. This caused an early model to hallucinate details from the paper's fictional scenarios, forcing Anthropic to implement unique mitigations. Speculatively, this kind of misalignment data could degrade the alignment of any models trained thereafter.^[1]

However, this result wouldn't rule out the hypothesis that the alignment-faking transcripts degraded Claude's alignment before they applied mitigations.

Data scraping practices are a serious problem. The tool we are currently releasing will not stop state-of-the-art actors. Since I wanted to at least mitigate the problem, I put out a bounty for a simple, open-source tool to harden data against scraping. The tool is now ready: easy-dataset-share. In less than 30 minutes and at a cost of $0, you can deploy a download portal with basic protections against scrapers, serving a canary-tagged dataset with modest protections against AI training.

Warning: easy-dataset-share will not stop sophisticated scrapers

Sophisticated scraping operations can bypass Cloudflare Turnstile for about $0.001 cents per trial (via e.g. CapSolver). The robots.txt and Terms of Service are not technically binding and rely on the good faith of the user, although the ToS does provide limited legal deterrence. Canary strings can be stripped from documents. Overall, this tool is just a first step towards mitigating dataset contamination. We later discuss improvements which might protect against sophisticated actors.

A download portal in minutes

We reduce the friction of serving data in a scraper-resistant fashion.

At most, users click a box. They don't have to complete any annoying tasks.

The protected download page after verification. — Generally, you'd just download a single .zip – possibly created by our `easy-datas`

...

Posts

Wikitag Contributions

Comments