December 18, 2024 Published by Praveen Ravichandran, Marco De Nadai, Divita Vohra, Sandeep Ghael, Manizeh Khan, Paul Bennett, Tony Jebara, Mounia Lalmas-Roelleke
Introduction
Generative Artificial Intelligence (AI), notably Large Language Models (LLMs), gives new alternatives to attach listeners with audio by means of deeply personalised and contextually related experiences. By integrating huge world information and adapting to numerous contexts, LLMs present a novel alternative to craft what we name private narratives—tales that resonate with listeners and assist familiarise them with suggestions on a person degree. This kind of engagement helps one among Spotify’s key objectives of connecting creators and listeners meaningfully. By leveraging the ability of LLMs alongside our experience in music, podcasts, and audiobooks, we intention to ship tailored experiences that assist listeners uncover new artists, creators, and authors and supply deeper context to their suggestions.
Focusing on contextualized suggestions is only one facet of Spotify’s broader AI journey. As we proceed to innovate with cutting-edge applied sciences, we’re unearthing a set of numerous purposes to redefine how customers work together with audio. These embrace 1) search and advice programs for extremely tailor-made content material discovery, 2) superior audio processing that powers options just like the pure, partaking voices of our English and Spanish AI DJs, and three) generative content material, resembling podcast chapterization, making podcasts simpler to navigate.
In this weblog put up, we discover two use circumstances that illustrate how Spotify makes use of LLMs to craft contextualized suggestions by means of personalised narratives: serving to customers uncover new artists with concise, significant explanations and delivering partaking, real-time insights tailor-made to every listener through AI commentary. Along the best way, we share insights into how we adapt open-source LLM fashions to satisfy our distinctive wants, guaranteeing these improvements are scalable, protected, and impactful.
Building a Backbone for Scalable AI Innovation
Over the previous decade, Spotify has leveraged numerous ML methods, together with graph neural networks, reinforcement studying, and extra, to remodel how customers have interaction with our intensive catalog. As we deepen our efforts in LLM improvement, it’s important to undertake a principled strategy to their integration, guaranteeing they’re aligned to ship contextualized suggestions by means of personalised narratives that resonate with each audiences and creators.
A sturdy spine mannequin is one constructing block to realize this, because it gives the pliability for fast experimentation and permits focused improvement. We can optimize a general-purpose LLM to suit exact product necessities whereas driving scalable innovation in personalised storytelling by leveraging a general-purpose LLM as a spine mannequin. Desirable traits of a powerful spine mannequin embrace:
- Broad World Knowledge: The spine mannequin ought to cowl a variety of basic and domain-specific information, making it well-suited to our numerous catalog of music, podcasts, and audiobooks. This information base permits it to craft contextual suggestions with out intensive retraining.
- Functional Versatility: An excellent spine mannequin ought to ideally excel at duties like operate calling and content material understanding, resembling matter extraction and security classification. This versatility permits speedy iteration on options that ship personalised and interesting person experiences.
- Community Support: Strong neighborhood help is essential for simplifying fine-tuning, providing environment friendly large-scale coaching and inference instruments, and driving steady enhancements. This helps hold us on the forefront of LLM developments, whereas enhancing our means to ship personalised suggestions.
- AI Safety: Safety is vital for guaranteeing a optimistic person expertise, notably within the context of personalised narratives. A spine mannequin should embrace safeguards to responsibly deal with delicate content material, forestall dangerous outputs, and guarantee compliance with regulatory requirements.
As a part of our analysis and exploration course of, we evaluated a variety of state-of-the-art fashions to establish these finest fitted to crafting contextualized and significant person experiences. While we proceed to make use of and refine a portfolio of fashions throughout our R&D groups, Meta’s household of Llama fashions emerged as a powerful candidate to assist us obtain essential components of this work, because it meets important standards for a dependable spine mannequin and is well-suited for area adaptation.
Use Cases
Traditionally, Spotify customers depend on cowl artwork and familiarity with an artist or style when deciding whether or not to interact with music suggestions. We have been exploring methods so as to add extra transparency and context to our suggestions to boost person confidence and encourage deeper engagement.
LLMs have proven important potential by providing significant, personalised context—virtually like a buddy’s advice would sound. This extra data will increase the probability that customers will discover new content material. By leveraging area adaptation, we’ve achieved promising outcomes that help this strategy. In the next sections, we delve into how this innovation transforms our platform, from AI-generated explanations for suggestions to personalised commentary from our AI DJ.
Contextualized Recommendations for New Releases
LLMs deliver a brand new dimension to Spotify’s personalization work by providing the power to elucidate why a specific merchandise would possibly resonate with customers. These explanations intention to assist customers perceive the rationale behind suggestions and to supply deeper perception into the content material.
That is why we explored how LLMs can generate concise explanations that add precious context to suggestions for music, podcasts, and audiobooks over the previous months. Combining the broad information of a spine mannequin with our experience in audio content material, we created explanations that ship personalised insights into really helpful content material. These explanations intention to spark curiosity and improve discovery. For instance, “Dead Rabbitts’ latest single is a metalcore adrenaline rush!” or “Relive U2’s iconic 1993 Dublin concert with ZOO TV Live EP.” By including this additional layer of context, advice explanations inform and encourage customers to discover extra of what Spotify has to supply.
Using LLMs to deliver the advice explanations product to life presents distinctive challenges, together with guaranteeing a constant technology fashion that aligns with our expectations, implementing sturdy security measures to stop inappropriate or dangerous outputs, mitigating hallucinations to keep away from inaccuracies, and successfully understanding person preferences to ship tailor-made significant explanations. Our preliminary assessments with zero-shot and few-shot prompting of open-source fashions resembling Llama highlighted the necessity for cautious LLM alignment, guaranteeing outputs are correct, contextually related, and cling to our requirements.
To obtain this, we adopted a human-in-the-loop strategy. Expert editors offered “golden examples” of contextualization. They additionally offered ongoing suggestions to handle challenges in LLM output, together with artist attribution errors, tone inconsistencies, and factual inaccuracies. In addition to steady human suggestions, we additionally included focused immediate engineering, instruction tuning, and scenario-based adversarial testing to generate the advice explanations. This iterative course of improved the general high quality of our advice explanations and aligned them extra intently with person preferences and creator expectations.
Our on-line assessments revealed that explanations containing significant particulars in regards to the artist or music led to considerably larger person engagement. In some circumstances, customers had been as much as 4 instances extra prone to click on on suggestions accompanied by explanations, particularly for extra area of interest content material.
These developments spotlight how we leverage state-of-the-art AI to redefine the chances of personalised discovery and person engagement. By delivering explanations and insights round suggestions that genuinely resonate with our customers, we improve their expertise and foster deeper engagement. At the identical time, we drive artist discovery by introducing listeners to new artists and content material they may not have explored in any other case.
Real-time Commentary for AI DJ
Another instance of how contextual suggestions present a deeper reference to creators is Spotify’s AJ DJ. Launched in 2023, DJ is a customized AI information that deeply understands listeners’ music tastes, offering tailor-made tune alternatives and insightful commentary on the artists and tracks it recommends. LLMs present a novel alternative to scale these personalised narratives, guaranteeing each listener receives wealthy, context-driven commentary that deepens their connection to the music and creators.
One key problem for LLM-based AI DJ commentary is attaining a deep cultural understanding that aligns with every listener’s tastes. At Spotify, music editors play a central function in assembly this problem, leveraging their style experience and cultural perception to craft experiences that really embrace the richness of worldwide music. By equipping these editors with generative AI instruments, we will scale their experience extra successfully than ever, enabling us to refine the mannequin’s output to make sure cultural relevance.
We constantly analysis and refine fashions powering our generative AI instruments to make sure they ship high-quality DJ commentaries. Through intensive comparisons of exterior and open-source fashions, we discovered that fine-tuning smaller Llama fashions produces culturally-aware and interesting narratives on par with state-of-the-art, whereas considerably decreasing prices and latency for this job.
Our personalised narratives for AI DJ permit listeners to discover new music and the tales behind the songs, deepening their connection to the content material. Similar to the contextual explanations instance, as we expanded the beta launch throughout a choose variety of markets, we discovered by listening to the commentary alongside private music suggestions, listeners are extra prepared to take heed to a tune they could in any other case skip. This transformative strategy redefines how customers have interaction with music and units the stage for additional improvements in personalised storytelling throughout our platform.
Adapting LLMs at Scale
Building personalised narratives with the assistance of spine fashions requires scalable infrastructure. To meet this want, we developed a complete information curation and coaching ecosystem that permits the speedy scaling of LLMs. This setup facilitates seamless integration of the most recent fashions whereas fostering collaboration throughout a number of groups at Spotify. These groups introduced their experience in constructing high-quality datasets, enhancing job efficiency, and guaranteeing the accountable use of AI. The curated datasets had been used for prolonged pre-training, supervised instruction fine-tuning, reinforcement studying from human suggestions, direct desire optimization, and thorough analysis.
LLMs like Llama are highly effective, general-purpose fashions, doubtlessly enabling a single mannequin to energy a number of use circumstances, resembling contextualizing suggestions and offering personalised AI DJ commentary. To make LLMs extra centered on Spotify wants, we tailored spine LLMs on a rigorously curated coaching dataset that included inside examples, information created by music consultants, and artificial information generated by means of intensive immediate engineering and utilizing zero-shot inference of state-of-the-art LLMs.
However, the use circumstances described on this doc are usually not the one ones Spotify focuses on. We additionally recognized a rising set of duties throughout the Spotify area the place AI may considerably improve efficiency. To discover this potential, we evaluated a variety of LLMs from 1B to 8B -sized fashions, benchmarking their zero-shot efficiency towards present non-generative, task-specific options. Llama 3.1 8B demonstrated aggressive efficiency. Building on this end result, we applied a multi-task adaptation of Llama, concentrating on 10 Spotify-specific duties. This strategy aimed to spice up job efficiency whereas preserving the mannequin’s basic capabilities. Throughout this course of, we used the Massive Multitask Language Understanding (MMLU) benchmark as a guardrail to make sure that the mannequin’s general talents remained intact.
Our outcomes demonstrated that the tailor-made adaptation led to important enhancements (as much as 14%) in Spotify-specific duties in comparison with out-of-the-box Llama efficiency. Additionally, we efficiently preserved Llama’s unique capabilities, with solely minimal variations in MMLU scores from the zero-shot baseline. This achievement underscores the potential for domain-adapting generative LLMs to spice up particular efficiency whereas retaining the mannequin’s sturdy foundational capabilities.
Distributed coaching is crucial to satisfy the excessive computational calls for of coaching LLMs with billions of parameters. A generally neglected facet of this course of is resilience to system failures or interruptions throughout prolonged, large-scale coaching phases on multi-node, multi-GPU clusters. To handle this, we developed a high-throughput checkpointing pipeline that asynchronously saves mannequin progress. By optimizing learn/write throughput, we considerably decreased checkpointing time and maximized GPU utilization.
Our LLM journey, nevertheless, extends far past fine-tuning. We are tackling challenges throughout the complete lifecycle, together with environment friendly serving and inference for offline and on-line use circumstances. We leverage light-weight fashions and superior optimization methods resembling immediate caching and quantization to realize environment friendly deployment, minimizing latency whereas maximizing throughput with out sacrificing accuracy. Open-source fashions additional improve our efforts by fostering a dynamic neighborhood of builders who contribute new instruments, deployment libraries, and analysis strategies, together with coaching accelerators and inference optimizations.
Integrating vLLM, a well-liked inference and serving engine for LLMs, has been a game-changer, delivering important serving efficiencies and decreasing the necessity for customized methods. vLLM permits low latency and excessive throughput throughout inference, permitting us to ship real-time generative AI options to thousands and thousands of customers. This versatile engine has additionally facilitated seamless integration of cutting-edge fashions like Llama 3.1, together with the 405B mannequin, instantly upon launch. This functionality empowers us to benchmark the most recent applied sciences and harness very giant fashions for purposes resembling artificial information technology.
We constructed a powerful basis for domain-adapted LLMs by means of collaboration and technological innovation. Our ongoing analysis and experiments in generative AI are unlocking new potentialities for person experiences. By working with trade leaders and open-source communities, we drive developments within the AI ecosystem and sit up for persevering with this journey with the broader trade.
Final Thoughts
Spotify’s strategy highlights the transformative potential of mixing cutting-edge generative AI with deep area experience to ship contextualized suggestions by means of personalised narratives. The assessments and learnings shared on this weblog showcase how LLMs may be tailored to push the boundaries of recommender programs, enabling real-time, extremely personalised experiences like AI DJ commentary and advice explanations. These improvements improve person engagement and redefine how creators and listeners join by means of context-driven storytelling.
Importantly, we’re dedicated to advancing this know-how past crafting new person experiences. By collaborating with trade leaders and open-source communities, we’re driving progress within the broader AI ecosystem whereas addressing vital challenges like infrastructure optimization and scalability. From foundational analysis to sensible developments like checkpointing pipelines and multi-task fine-tuning, we’re constructing options that ship distinctive effectivity and effectiveness at scale.
By sharing concrete purposes, we intention to reimagine the potential of this know-how and contribute to a broader dialog about the way forward for generative AI in recommender programs. We intention to encourage new methods of delivering significant, personalised experiences that foster deeper connections between customers and content material.