360Brew: LLM-based Personalized Ranking and Recommendation
summary
Claims a 150B-parameter foundation model called Blue XL is trained with promptification and instruction fine-tuning and then distilled stepwise to usable sizes (down to ~3B) using gradual pruning and mixed-precision (FP8) quantization. It scores up to 500 candidates without cross-attention between items, improves cold-start performance versus production models as interactions drop, generalizes on 4+ unseen tasks and is validated on a 50–60 benchmark suite, and yields up to 7x latency reduction and 30x GPU throughput increase across releases.
AI-written summary · about
chapters (9)
- 0:00Introduction to 360Brew
- 1:46Mission for a single model
- 3:35Building the model with promptification
- 5:00Training pipeline and model scaling
- 6:46Leveraging data and context length
- 8:53Results on cold start users
- 9:59Generalization to new domains
- 10:40Serving large models efficiently
- 15:58Q&A session
related talks
41mBoosting Ranking Performance with Minimal Supervision
01h 38mTraining Large Language Models: Practices and Research Questions
40mScaling generative AI workloads with efficient model choice
41mMeet DBRX, the New Standard for High-quality LLMs, and How...
33mBuild Better AI with First-Party Data and Fine-Tuning
20mEverything I Learned Training Frontier Small Models
30mLLM-Based Recommendation Systems
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures