GPT-4o’s Chinese language token-training knowledge is polluted by spam and porn web sites

The brand new tokenizer has 200,000 tokens in whole, and about 25% are in non-English languages, says Deedy Das, an AI investor at Menlo Ventures. He used language filters to rely the variety of tokens in several languages, and the highest languages, in addition to English, are Russian, Arabic, and Vietnamese.

“So the tokenizer’s predominant affect, in my view, is you get the price down in these languages, not that the standard in these languages goes dramatically up,” Das says. When an LLM has higher and longer tokens in non-English languages, it will possibly analyze the prompts sooner and cost customers much less for a similar reply. With the brand new tokenizer, “you’re taking a look at nearly 4 occasions price discount,” he says.

Das, who additionally speaks Hindi and Bengali, took a have a look at the longest tokens in these languages. The tokens replicate discussions occurring in these languages, so that they embody phrases like “Narendra” or “Pakistan,” however frequent English phrases like “Prime Minister,” “college,” and “worldwide” additionally come up incessantly. In addition they don’t exhibit the problems surrounding the Chinese language tokens.

That probably displays the coaching knowledge in these languages, Das says: “My working concept is the web sites in Hindi and Bengali are very rudimentary. It’s like [mostly] information articles. So I’d anticipate this to be the case. There aren’t many spam bots and porn web sites attempting to occur in these languages. It’s largely going to be in English.”

Polluted knowledge and an absence of cleansing

Nevertheless, issues are drastically completely different in Chinese language. In accordance with a number of researchers who’ve seemed into the brand new library of tokens used for GPT-4o, the longest tokens in Chinese language are nearly solely spam phrases utilized in pornography, playing, and scamming contexts. Even shorter tokens, like three-character-long Chinese language phrases, replicate these subjects to a major diploma.

“The issue is evident: the corpus used to coach [the tokenizer] just isn’t clear. The English tokens appear tremendous, however the Chinese language ones aren’t,” says Cai from Princeton College. It isn’t uncommon for a language mannequin to crawl spam when amassing coaching knowledge, however normally there shall be vital effort taken to scrub up the information earlier than it’s used. “It’s attainable that they didn’t do correct knowledge clearing with regards to Chinese language,” he says.

The content material of those Chinese language tokens might recommend that they’ve been polluted by a selected phenomenon: web sites hijacking unrelated content material in Chinese language or different languages to spice up spam messages. 

These messages are sometimes commercials for pornography movies and playing web sites. They may very well be actual companies or merely scams. And the language is inserted into content material farm web sites or generally professional web sites to allow them to be listed by search engines like google, circumvent the spam filters, and are available up in random searches. For instance, Google listed one search end result web page on a US Nationwide Institutes of Well being web site, which lists a porn web site in Chinese language. The identical web site identify additionally appeared in no less than 5 Chinese language tokens in GPT-4o.