Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify

AI Engineer · 19 min · 145 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Welcome everyone.
00:13Um really glad and thank you Deven for inviting us as well to talk today.
00:18Uh so um uh Jackie and I are going to present today about making LLMs speak Spotify and how we turned our recommendation system on its head to be LLM native.
00:31Uh just a few quick words about myself.
00:33Uh I joined Spotify a year ago.
00:35Before that I was at Google working on personalization for Google search and prior to that I was uh working on personalization at Netflix.
00:44Um all right so let's dive in.
00:47Um first a couple of numbers to describe the scale of the problem that we have to solve.
00:54Um Spotify has about I was a little bit sad not to see it on the DAU by chart because it should be up there but it has about 760
01:02million uh active uh users monthly uh in about 184 markets.
01:08But one thing that makes uh the Spotify personalization problem particularly challenging is the size of its catalog.
01:15Uh so of course you know Spotify has basically all music ever published a bit more than 100 million uh music tracks but it also has a range of videos, podcasts and audio books as well.
01:26So the matching problem uh is uh actually surprisingly uh complicated.
01:32So today what we're going to talk about Jackie and I is basically the extent of this matching problem and how we uh how we are solving it.
01:40I'm going to talk about a little bit of history, the new phase that we're entering and then Jackie is going to help us uh go into the guts of the modes as well to understand how those things are trained as well.
01:50Uh so a little bit on history.
01:52Uh initially Spotify personalization was really built around curation.
01:57So uh people basically manually assembling playlists that target specific taste.
02:02And that's still a very common uh use case on Spotify.
02:05Right now we have about 10 billion playlists with many many uh created every every hour.
02:12Uh then you know Spotify moved into taking these creation signals and order signals and moving into recommendations.
02:19So basically being able to turn these uh creation signals to something that can be applied at scale.
02:25Uh and a great example of that would be discover weekly for example that was launched in 2014 as one of the one of the early like recommendation use case on on Spotify.
02:36But the phase that we entering now which uh uh which we're going to talk about in more details is something that we call generative personalization where we are not only solving a matching problem from the user to the content.
02:47We're also solving the ability to generate an experience that is interactively and dynamically shaped around each user.
02:58And that transition from recommendations to generative personalization involves a couple of different uh a couple of different big shifts.
03:07One is move from moving moving from personalization as guessing uh where basically you have a ranking algorithms that spits out a brand ranked list of entities in your catalog
03:18to personalization as reasoning that can introspect these results and really try to understand whether that's indeed the right match for this user in this particular context.
03:28The other aspect as well is moving from blackbox algorithms.
03:32uh it's very difficult to fully introspect a multi-stage ranking system for example to transparent and steerable personalization where the user is always fully in control.
03:42So uh basically giving the ability to these mods to speak and understand English.
03:50The other thing that these these systems can do is uh uh not stopping just at recommending but also generation of experiences and explaining as well.
04:00uh and we're going to show a couple of examples of that.
04:05So one example that uh we launched a couple of years ago uh so pretty early in that journey was the Spotify DJ.
04:12Uh so the Spotify DJ is something that you can spin up that will start playing music for you of course personalized uh and but one interesting thing about it is that since last year you can tap that button on the bottom uh right and
04:26steer it in whatever direction you see fit.
04:28So at any point you can chime in and let the algorithm know what you want and it's going to steer the session in the in in that in that direction.
04:37Uh another example of what we call generative personalization is uh showcased in a prompted playlist here that Deanch showed a little bit earlier as well.
04:47Uh and here you can you basically have full unfettered access to the recommendation algorithm that Spotify has.
04:54uh and you can uh prompt it with very high level prompts or very detailed prompts.
05:00On the left hand side here, I have a prompt that tells me uh create me a playlist of bands that are playing in San Francisco tonight.
05:06Uh so turns out there's a bunch of good shows if you're excited to check them out.
05:11On the right hand side, you see a prompt that is asking for a playlist to accompany me on my run.
05:15And what's interesting with the right hand side as well is that you will see that the experience itself gets dynamically shaped as a function of the request on the user to be able to introduce itself and different phases in my run as well.
05:29Uh, another one that I'm really excited about, we launched it in New Zealand a couple of months ago and uh, it's coming soon in more markets is something called the taste profile.
05:39And that basically gives you the ability to introspect in natural language what the Spotify algorithm has understood about you in a way that you can edit and refine.
05:49Uh, so if you see something that's missing or something that's wrong.
05:52Uh so for example one of my edit is that all Disney music are my kids because we have a bunch of shared devices at home but please don't recommend that to me that's not my taste please.
06:02Uh and the algorithm would then take that into account uh making sure that we never recommend this this in the wrong context.
06:09Uh and similarly you can also use the taste profile to share some more aspiration uh goals as well.
06:14Uh so getting into a new genre, getting into a new topic, uh learning about a new language for example, all of these things can be can be done.
06:23Uh another one that's coming uh soon uh which we uh announced very recently is something called personal podcast where the the the generative personalization system doesn't stop at just recommending and ascending experiences but also generating content as well.
06:37In this particular example, I'm generating a daily brief that's uh that's generated on a cadence daily.
06:43Uh and I'm going to uh let it uh tell me about what's happening in my community.
06:51All right.
06:51So now to go into the guts of it.
06:54So there's one big system that controls like all of these different applications I mentioned and we call that internally some the large taste model.
07:01We're not great at naming these internal things.
07:04Um so it has a couple of properties.
07:06One is that it understands every historical interaction piece of content on Spotify.
07:10Uh it combines prediction and reasoning to the point that I mentioned earlier.
07:14So not only guessing but also reasoning layered on top and it gives users the ability to shape and generate experiences in real time.
07:21So it's fully steable and promptable by uh by users.
07:25Um and uh as of today about one in four US premium subscribers interact with that system on a daily basis as well.
07:32So that's pretty exciting.
07:33Um, one thing that's exciting as well to uh is that deploying this system across existing recommendation surfaces as well also led to some gains.
07:42Uh, we saw gains on autoplay, we saw gains on podcast discoveries.
07:47Uh, we saw gains on users interacting with DJ messages as well.
07:50So, we saw pretty pretty sizable gains across the board as well by deploying the system.
07:55On that note, I'm going to hand it over to Jackie to talk to us about what's one of the core component that underpins this whole system.
08:03Hi everyone, I'm Jackie or Jacqueline, a staff machine learning engineer at Spotify.
08:09So let's dive a little bit deeper and talk about how these models are actually trained at least at Spotify.
08:16So semantic IDs were presented in the previous talk, but that is how we are embedding these openweight LLMs with knowledge of Spotify's catalog.
08:27They are created by taking existing content embeddings such as podcast episode embeddings and applying a quantization algorithm to convert them to a set of discrete tokens.
08:42We then take a openweight LLM such as Quen and we modify its vocabulary to add these new special tokens and then we fine-tune the model to be able to understand
08:54both natural language as well as these new special tokens semantic IDs that represent Spotify catalog entities.
09:03So we can power experiences such as this where the user can ask in natural language for a podcast on morality.
09:11And that is passed to the prompt along with their listening history represented as semantic IDs.
09:18And the model responds both with a relevant semantic ID podcast episode as well as a natural language description of why they recommended that to this user.
09:33So how is this model actually trained?
09:35Um we published a paper linked here uh describing our training paradigm called NEO which consists of four distinct stages.
09:47The first I already covered which is the semantic foundation stage where we construct meaningful semantic ID tokens and then add them to an openway LLM's vocabulary.
10:01The second stage we call domain grounding in which we align these new semantic ID token embeddings in the original language embedding space.
10:13We do this by learning a birectional mapping between semantic ids to text, text to semantic ids and any combination.
10:26And we actually freeze the LLM backbone at this stage and only train the new semantic ID embeddings.
10:35So the original model weights and embeddings are frozen and we just learn those new semantic ID tokens and this helps us to mitigate catastrophic forgetting of the pre-trained LLM's
10:49core language abilities.
10:52The third stage we call capability induction which is multitask instruction tuning on tasks that Spotify cares about such as the ones shown here.
11:03next item recommendation retrieval etc. This is done by unfreezing the whole model all of its weights and embeddings and running either full parameter fine-tuning or Laura fine-tuning on the multiple Spotify tasks
11:21and then there is an optional fourth stage to do post- training such as RL fine-tuning etc. So, how much of a difference does this four-stage training paradigm actually make?
11:36I'll dive into a few of the abilations we've done to investigate this.
11:42First, we assessed whether multitask training is actually hurting performance by comparing the multitask model against single task variance.
11:52And we consistently saw that across our tasks, the multitask model can match or actually beat the single task performance indicating that there is some positive cross-learning happening across the tasks.
12:06This is particularly noticeable for audiobook recommendations.
12:11If you look here, which is a new newer content type at Spotify, demonstrating that these multitask models can help with cold start entities by learning from other items in the catalog such as podcast recommendations,
12:28how to make meaningful audiobook recommendations.
12:33Then we did some abilations on the actual training recipe.
12:36We evaluated both dropping the frozen backbone domain grounding stage altogether as well as combining the domain grounding stage with the capability induction stage in a multitask instruction tuning stage
12:53that those are rows A and B here in the middle and we saw for both of those that it degraded performance but actually the biggest drop in performance was from
13:03using a randomly initialized backbone instead of the pre-trained openweight LLM that we are using.
13:11We also evaluated using continuous pre-training for the domain grounding stage.
13:17And as you can see, it's it's minimal actual degradation on the task specific performance.
13:23But where continuous pre-training really hits us is on the natural language and world knowledge capabilities of the pre-trained backbone LLM we are using.
13:34after we do continuous pre-training, it goes to essentially zero versus if we do the frozen backbone domain grounding.
13:41We retain all of that core language ability and are still able to learn the semantic ids.
13:47I want to call out that these abilations were done with Quen, but we also validated that these findings hold with Llama.
13:55So, it is not specific to the model backbone, but actually the training paradigm itself.
14:03And then lastly, we did some investigation on different inference strategies and their effect on accuracy versus latency.
14:13We tested beam search with both constrained decoding and not.
14:18And we saw that even without constrained decoding, we can generate valid semantic IDs 98% of the time.
14:27Constrained decoding does add a little latency overhead, but it's also helpful for specific cases where you want to target specific types of content, such as only make new content recommendations,
14:40for example.
14:42We also compared top P sampling to beam search and saw that top P sampling pretty significantly hurts our accuracy.
14:50So although beam search is a little more latency intensive, we decided that trade-off worked for us.
15:02So what is meaningful about this?
15:04There have been lots of work in the industry in the space on generative semantic ID retrieval, toolbased LLM recommenders, the plum paper, etc. But NEO is actually the first example of combining all these capabilities into one system
15:21that understands grounded catalog items, is naturally language steerable, can do search, recommendation, explanation use cases, as well as low latency tool-free inference at an industrial scale.
15:41And we are using this in production today.
15:45Um there's a paper linked as well here for how we're using this to power podcast discovery.
15:53What we saw is that using a model train like this, we can break users out of their habitual patterns and get them to listen to more unfamiliar content.
16:04And we saw huge wins online with this.
16:10So none of this works without meaningful evaluations.
16:14So I want to talk about that a little bit as our recommendations are becoming more generative and explanatory.
16:23We our original maybe traditional offline eval metrics are not sufficient.
16:32Yes, they can tell us whether or not the user interacted with that content, but they can't tell us whether or not the user or that recommendation makes sense for the user,
16:44whether the explanation is accurate, whether it aligns with the user's intent, etc. So, that's where LLM judges really shine.
16:52But in our work at Spotify, we really find that you need to invest in grounding your LLM judges in meaningful data.
17:01so that they can be reliable evaluators that align with human preferences.
17:07So, a few examples um for evaluating the podcast recommendations that I just mentioned, we create textual user profiles summarizing the users's listening history and that is passed to the LLM as a judge and we saw that this corresponded with a 75%
17:27alignment between the LLM judge and human preferences.
17:32Similarly, you can use actual behavioral signals to ground these LLM judges.
17:37So, for example, for a search task, you can for a given query, you can take similar queries and how the user has interacted with them in the past and pass that to the model.
17:50And we saw that overall it increased alignment by 5% but on ambiguous queries it actually increased alignment by 91%.
18:01Showcasing the value these this grounding of the LLM judge plays especially in ambiguous cases where LLM judges tend to struggle.
18:13Lastly, we used grounded LM judges to scale up our Cranfield style collections.
18:21These are evaluation sets that are constructed by taking candidates from multiple different sources, creating a pool, and then using a human to rank that pool.
18:33However, that human ranking stage is expensive.
18:37So, we invested in significant grounding for our LM as judge and are able to have an LLM judge that aligns with uh human system rankings with an agreement value of 0.87.
18:56In summary, um like we're going to hear a lot about today in all the talks, there is a new era of personalization among us.
19:05this generative personalization.
19:07And if you want to power language steerable personalized recommendations for your user, this is how we taught openweight LLMs to speak Spotify.
19:20Thank you.
19:22[applause]
19:37[music]