🚀 Releasing Amazon Reviews 2023 dataset! With *500+M* user reviews, *48+M* items, *60+B* tokens, all from 33 categories, Amazon Reviews, one of the largest, most widely-used review dataset has come to its fourth generation. A thread 🧵 https://t.co/e6bq7mK3NJ

Q: Can we pre-train LLMs efficiently (and better?) via data pruning? A: Yes! Q: How? A: (secret) Prompt LLMs for data quality 🤫 Check out our latest work @GoogleDeepMind - “How to Train Data-Efficient LLMs” 📖 https://t.co/0Lc6WDQIpm An expensive thread 🧵(RTs appreciated!)

Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long

Fine-grained control/editing in text-to-music diffusion models w/NO TRAINING? Presenting DITTO: Diffusion Inference-Time T-Optimization for Music Generation 📖: https://t.co/JjXxWkl3wW 🎹: https://t.co/CqLtgaZoBC w/@McAuleyLabUCSD @BergKirkpatrick @NicholasJBryan🧵

Our paper on interesting findings about “LLMs & RecSys” has just been accepted as a full paper in #ecir2024 The most delightful thing is that we got really high-quality, detailed, and constructive reviews. Thanks reviewers from @ecir2024 ! A thread 🧵 https://t.co/HVMBAr4iX6

Lead sheets concisely describe music, but can we improve their compressive ability w.r.t. the original score? Check out our new work - Unsupervised Lead Sheet Generation via Semantic Compression 📖 https://t.co/lknNqygUMn w/@NikitaSrivatsan @BergKirkpatrick @McAuleyLabUCSD 1/n

Farzi Data: Autoregressive Data Distillation paper page: https://t.co/Hc1urSUXsO study data distillation for auto-regressive machine learning tasks, where the input and output have a strict left-to-right causal structure. More specifically, we propose Farzi, which summarizes an

Wonderful #KDD2023 🥳 (with a cute Thai Tea🥤)

🤖️ Are LLMs good Conversational Recommender Systems (CRS) ? We (@McAuleyLabUCSD and @NetflixResearch) let LLMs generate movie names directly in response to natural-language user requests. Key observations in the experiments:

Highly grateful! Definitely recommend the streamlined publication experience @TmlrOrg For people intersted in data distillation, do checkout our survey - it designed to be to-the-point, and does not require a lot of prerequisite knowledge. Any feedback is highly appreciated!

Ecstatic to join @DeepMind as a research intern for the summer -- looking forward to new friends and being surrounded by the smartest of smartest 🦾 Please DM me if you're around MTV-CE, let's go for a coffee ☕️

Featuring the awesome work of @JulianMcauley @XuCanwen Zexue He, Zhankui He ⬇️

Researchers @UCSanDiego developed algorithms to rid speech generated by online bots of offensive language, on social media and elsewhere. For more stories about the #UCEngineer impact on #CyberSecurity, visit https://t.co/zZ8XanVkQW #Eweek2023 #UCEngineer

Announcing with shaky hands and much delight: Our "conversational critiquing" paper is selected for "Highlights of ACM RecSys '22". 🎉 Didn't know before what it is like to be among the bests of a conf ~ @ACMRecSys Paper: https://t.co/yyD6yHPcTM @ShuyangLi2 @McAuleyLabUCSD 🤩

Conventional #recsys wisdom: "better to go wide than deep". Our paper: go infinitely-wide, compute the solution in closed-form with a single hyper-parameter, and considerably beat all SoTA. Furthermore, can you get the same performance with just 500 fake users? Yes! A thread 🧵

Happy to share that our paper "Deep Performer: Score-to-Audio Music Performance Synthesis" has been accepted to @ieeeICASSP 2022! 🥳 This joint work with @CongZhou1, @BergKirkpatrick and Julian McAuley (@McAuleyLabUCSD) is based on my internship work at @Dolby last summer. 🎶