Fine-tuning large language models (LLMs) on a mixture of diverse datasets poses challenges due to data imbalance and heterogeneity. Existing methods often address these issues across datasets (globally) but overlook the imbalance and heterogeneity within individual datasets (locally), which limits their effectiveness. We introduce Hierarchical Balancing Optimization (HBO), a novel method that enables LLMs to autonomously adjust data allocation during fine-tuning both across datasets (globally) and within each individual dataset (locally). HBO employs a bilevel optimization strategy with two types of actors: a Global Actor, which balances data sampling across different subsets of the training mixture, and several Local Actors, which optimizes data usage within each subset based on difficulty levels. These actors are guided by reward functions derived from the LLM's training state, which measure learning progress and relative performance improvement. We evaluate HBO on three LLM backbones across nine diverse tasks in multilingual and multitask setups. Results show that HBO consistently outperforms existing baselines, achieving significant accuracy gains. Our in-depth analysis further demonstrates that both the global actor and local actors of HBO effectively adjust data usage during fine-tuning. HBO provides a comprehensive solution to the challenges of data imbalance and heterogeneity in LLM fine-tuning, enabling more effective training across diverse datasets.

By pretty much every objective measure, HBO's adaptation of the hit postapocalyptic video game The Last of Us has been a roaring success. Never before has a video game narrative been molded into Emmy nominations and such warm reception among respectable critics, industry darlings, and people who have no idea what the term "one-shotting" means. You'd think that the devotees who first fell in love with the game back when it was originally released in 2013 would be toasting the cultural ascendance of their favorite medium--and especially how the story's complicated morality has impacted those who've never picked up a controller. And yet, for as long as the show has been on television, its most dogmatic fans have been caught up in a controversy of much inferior consequence: Specifically, they're furious that Bella Ramsey doesn't look much like Ellie. On the most basic level, this observation is correct.

"Fudge," I remember saying, only I didn't say fudge, I said fuck, a word for adults. I was playing The Last of Us, a narrative video game for adults about a zombie apocalypse, and I had just died for what seemed like the thousandth time in the first room with a "clicker," the game lore's name for a medium-difficulty enemy. These "infected"--it's classier not to call them zombies, and this is a classy zombie-combat game, one with a story--had become misshapen thanks to a cordyceps brain infection, which devoured mankind almost overnight. The clicker was ghastlier than others, because it had lived long enough for the infection to fully engulf its formerly human face, fungal fibers enrobing it, teeth jutting out like barbs. An older infected is a more resilient one.

"The Last of Us Part II" attempted to get players to empathize with NPC deaths more so than in the first game through dogs that mourned their masters' deaths or friends of the fallen calling out for them by name. The show continues to intensify death, including early moments during the pandemic's initial outbreak in the pilot, through the addition of tendrils. These thin, stemlike threads slowly grow from an infected's mouth. Upon contact with a human, they too become infected, offering an entirely new method of transmission. Characters can still be bitten by zombies, but spores -- which allowed for the game's disease to be airborne -- were ditched for this first season.

This content can also be viewed on the site it originates from. When the British actor Bob Hoskins agreed to star in "Super Mario Bros.," he had little sense of what he was getting into. The year was 1992, and, although the title on which the film was based had sold tens of millions of copies, a feature-length live-action adaptation of a video game had never been attempted. The movie's eventual tagline, "This ain't no game," reflected a self-conscious distance from its source material: a convoluted parallel-universe plot recast the heroes as Italian American handymen from Brooklyn and the princess they set out to save as an N.Y.U. Hoskins himself hadn't even heard of the Nintendo franchise--but when his kids learned that he would be playing Mario they excitedly showed him the game.

I love my Apple TV, but I can't stand the old Siri remote that came with it. Its touchpad is too sensitive, and I never know if it's in the right orientation when I grab it. Apple did fix all that with its second-generation Siri remote, but why buy another one when the one I currently have still works--especially when I can use a few handy Siri commands to do more than what the touchpad can do. On the Apple TV and accompanying Siri remote there's a handy microphone button that lets you chat directly with Siri to get your TV to do things like search for movies in a specific genre or year, ping other devices, and even fast forward by a specific amount of time. Not every command is intuitive, though, so it's helpful to know a few of them off-hand before you binge another series.

The recent introduction of thermodynamic integration techniques has provided a new framework for understanding and improving variational inference (VI). In this work, we present a careful analysis of the thermodynamic variational objective (TVO), bridging the gap between existing variational objectives and shedding new insights to advance the field. In particular, we elucidate how the TVO naturally connects the three key variational schemes, namely the importance-weighted VI, Renyi-VI, and MCMC-VI, which subsumes most VI objectives employed in practice. To explain the performance gap between theory and practice, we reveal how the pathological geometry of thermodynamic curves negatively affects TVO. By generalizing the integration path from the geometric mean to the weighted Holder mean, we extend the theory of TVO and identify new opportunities for improving VI. This motivates our new VI objectives, named the Holder bounds, which flatten the thermodynamic curves and promise to achieve a one-step approximation of the exact marginal log-likelihood. A comprehensive discussion on the choices of numerical estimators is provided. We present strong empirical evidence on both synthetic and real-world datasets to support our claims.

The three comedians talk about what it was like to craft a bilingual TV show with dialogue in both English and Spanish and why the show isn't set in a particular country. They also discuss the show's supernatural elements, which intentionally lack specific rules and logic. After the interview, June and co-host Isaac Butler help a listener who's feeling unproductive in her new workplace. Send your questions about creativity and any other feedback to working@slate.com.

HBO is working on a television adaptation of the bestselling video game The Last of Us. It will be written by Craig Mazin, the creator of highly acclaimed TV series Chernobyl, together with Neil Druckmann, the creative director on the game and its forthcoming sequel, according to the Hollywood Reporter. The series is set to follow the events of the original game, which takes place in a United States devastated by a deadly parasitic fungus that transforms infected victims into monsters. The action follows grizzled survivor Joel, who is hired to transport seemingly immune teenage girl Ellie out of a quarantine zone and across the country to a lab where a cure can be developed. Praised for its bleak, violent narrative, the game was a major hit on PlayStation, with a sequel due in May.

Sign in to report inappropriate content. VICE's Elle Reeve heads to China to investigate the rise of facial recognition technology -- and what that means for all of us. Check out VICE News for more: http://vicenews.com Follow VICE News here: Facebook: https://www.facebook.com/vicenews