Below the hood
Getting LLaMA 2 able to launch required loads of tweaking to make the mannequin safer and fewer more likely to spew poisonous falsehoods than its predecessor, Al-Dahle says.
Meta has loads of previous gaffes to be taught from. Its language mannequin for science, Galactica, was taken offline after solely three days, and its earlier LlaMA mannequin, which was meant just for analysis functions, was leaked on-line, sparking criticism from politicians who questioned whether or not Meta was taking correct account of the dangers related to AI language fashions, akin to disinformation and harassment.
To mitigate the danger of repeating these errors, Meta utilized a mixture of completely different machine studying strategies aimed toward bettering helpfulness and security.
Meta’s method to coaching LLaMA 2 had extra steps than standard for generative AI fashions, says Sasha Luccioni, a researcher at AI startup Hugging Face.
The mannequin was skilled on 40% extra information than its predecessor. Al-Dahle says there have been two sources of coaching information: information that was scraped on-line, and an information set fine-tuned and tweaked in accordance with suggestions from human annotators to behave in a extra fascinating means. The corporate says it didn’t use Meta person information in LLaMA 2, and excluded information from websites it knew had plenty of private data.
Regardless of that, LLaMA 2 nonetheless spews offensive, dangerous, and in any other case problematic language, similar to rival fashions. Meta says it didn’t take away poisonous information from the info set, as a result of leaving it in would possibly assist LLaMA 2 detect hate speech higher, and eradicating it may threat unintentionally filtering out some demographic teams.
However, Meta’s dedication to openness is thrilling, says Luccioni, as a result of it permits researchers like herself to check AI fashions’ biases, ethics, and effectivity correctly.
The truth that LLaMA 2 is an open-source mannequin can even enable exterior researchers and builders to probe it for safety flaws, which can make it safer than proprietary fashions, Al-Dahle says.
Liang agrees. “I am very excited to attempt issues out and I feel will probably be useful for the group,” he says.
