under the hood
Al-Dahle says that getting LLaMA 2 ready for launch required a lot of fine-tuning in order to make the model safer and reduce the likelihood of spreading toxic lies compared to its predecessor.
Meta has had its fair share of mishaps in the past. The language model Galactica, focused on scientific content, was taken offline just three days after launch, and the previous LlaMA model, which was only intended for research purposes, had its online version leaked, leading to criticism from politicians questioning whether Meta properly considered the associated risks of AI language models for misinformation and harassment.
To reduce the risk of these errors recurring, Meta has applied a blend of different machine learning techniques aimed at increasing helpfulness and safety.
Sasha Luccioni, a researcher at the artificial intelligence enterprise Hugging Face, says that Meta’s training approach for LLaMA 2 involves more steps than usual for productive AI models.
The model was trained with 40% more data than its predecessor. Al-Dahle explains that there were two sources of training data: scraped online data and a fine-tuned dataset based on feedback from human annotators to behave more desirably. The company says it did not use Meta user data in LLaMA 2 and excluded data from sites known to contain a large amount of personal information.
Nevertheless, just like rival models, LLaMA 2 still exhibits offensive, harmful, and otherwise problematic language. Meta says it did not remove toxic data from the dataset, as keeping it in can help LLaMA 2 better detect hate speech and removing it may carry the risk of inadvertently filtering certain demographic groups.
However, Luccioni finds Meta’s commitment to transparency exciting because it allows researchers like herself to examine the bias, ethics, and performance of AI models properly.
Al-Dahle states that LLaMA 2 being an open-source model will allow external researchers and developers to investigate it for security flaws, making it more secure than proprietary models.
Liang shares the same sentiment. “I’m very excited to try it out, and I believe it will be beneficial for society,” she says.