High -end legal case that the company Meta Targeting, revealing numerous domestic documes from the company. Among them, there is a docume that has attracted the atteion of some scholars of artificial ielligence and new details on how Teaching AI models It reveals like Llama.
According to Business Insider, the documes explained that meta researchers from a process called “AblationThey use to determine which data were effective in improving the performance of the LLAMA model. The term borrowed from the medical field refers to the process of deletion or deliberate destruction of a section to investigate its effect on the overall system performance.
In the field of artificial ielligence, obstruction means removing or replacing part of the system or data to determine the role of that part in the final performance.
In one of these experimes, Meta has part of her educational data with books from the database Libgen Replaceme that has been illegally released, then re -taught the LLAMA model to evaluate the effect of the replaceme.
In another experime, Meta added scieific, technology and fiction books to educational data, and in another experime, only fiction books were eered io the training process. According to iernal documes released in court, the performance of the LLAMA model in all two experimes has improved dramatically in industrial benchmarks.
Confideial results of meta abolition testing
Ablation tests meta First step Education focuses in which massive data is used to familiarize the model with the concepts of the real world.
In the first experime, the addition of the BooiQ benchmarking science, technology and fiction books 4.5 % Improveme. Also adding fiction books 6 % Has made recovery.
Peter Henderson, a professor of computer science at Princeton University, has published charts of meta’s iernal documes in X, which shows the results of these improvemes.
Booiq coains a set of about 16,000 yes/ no questions that the model must answer. The more the model answers the questions, the better its performance is assessed. The 5 % improveme, the meta model, has been able to answer about 800 more questions.
Common but confideial technique for artificial ielligence companies
Abullah has become a common way in Meta and other companies active in the field of artificial ielligence; For example, Insider says one of the Meta engineers in LinkedIn has announced that in the LLAMA 4 developme process and earlier versions, he has performed more than 100 abolition tests.
Meta, of course, does not publish the results of these experimes, and other companies are sile.
Nick Vince, an assista professor at the University of Simon Freezer’s Computer Science Faculty, says one of the possible reasons for this secrecy is that if the data is precisely improved, the original owners of the data can apply for fees.
She says:
“If these general numbers are announced, coe generating organizations may take a stronger legal position.”
Finally, Vince hopes that such disclosures on confideial meta tests will lead to a new system to attribute credit to educational data resources and fair financial compensation. She says:
“Artificial ielligence chats are based on the fact that a human being has done somewhere useful, written and published. This technology has re -packaged that information and hopes to make it more useful. Finally, everything goes back to humans. Without this data, artificial ielligence models will not perform well. “Documeation of abolition testing can help create healthy flow and enhance institutions that encourage coe and knowledge production.”
The report has been released as technology gias criticized meta for inappropriate regulation of court documes and sensitive information disclosure.
